WO2003081574A1 - Pattern recognition - Google Patents

Pattern recognition Download PDF

Info

Publication number
WO2003081574A1
WO2003081574A1 PCT/IB2002/000954 IB0200954W WO03081574A1 WO 2003081574 A1 WO2003081574 A1 WO 2003081574A1 IB 0200954 W IB0200954 W IB 0200954W WO 03081574 A1 WO03081574 A1 WO 03081574A1
Authority
WO
WIPO (PCT)
Prior art keywords
distortion measure
control signal
feature vector
subset
contributions
Prior art date
Application number
PCT/IB2002/000954
Other languages
French (fr)
Inventor
Imre Kiss
Marcel Vasilache
Original Assignee
Nokia Corporation
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Nokia Corporation filed Critical Nokia Corporation
Priority to PCT/IB2002/000954 priority Critical patent/WO2003081574A1/en
Priority to JP2003579211A priority patent/JP4295118B2/en
Priority to EP02722529A priority patent/EP1488408A1/en
Priority to KR1020047015018A priority patent/KR100760666B1/en
Priority to AU2002253416A priority patent/AU2002253416A1/en
Priority to CNB028286480A priority patent/CN1295672C/en
Priority to US10/402,367 priority patent/US7269556B2/en
Publication of WO2003081574A1 publication Critical patent/WO2003081574A1/en

Links

Classifications

    • GPHYSICS
    • G10MUSICAL INSTRUMENTS; ACOUSTICS
    • G10LSPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
    • G10L15/00Speech recognition
    • G10L15/28Constructional details of speech recognition systems
    • GPHYSICS
    • G10MUSICAL INSTRUMENTS; ACOUSTICS
    • G10LSPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
    • G10L15/00Speech recognition
    • G10L15/02Feature extraction for speech recognition; Selection of recognition unit
    • GPHYSICS
    • G10MUSICAL INSTRUMENTS; ACOUSTICS
    • G10LSPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
    • G10L15/00Speech recognition
    • G10L15/08Speech classification or search
    • G10L15/14Speech classification or search using statistical models, e.g. Hidden Markov Models [HMMs]
    • G10L15/142Hidden Markov Models [HMMs]
    • GPHYSICS
    • G10MUSICAL INSTRUMENTS; ACOUSTICS
    • G10LSPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
    • G10L15/00Speech recognition
    • G10L15/22Procedures used during a speech recognition process, e.g. man-machine dialogue
    • G10L2015/226Procedures used during a speech recognition process, e.g. man-machine dialogue using non-speech characteristics
    • G10L2015/228Procedures used during a speech recognition process, e.g. man-machine dialogue using non-speech characteristics of application context

Definitions

  • the present invention relates to pattern recognition, where a set of feature vectors is formed from digitized incoming signals, and compared with templates of candidate patterns.
  • incoming signals are digitized, and a sequence of feature vectors are formed. These feature vectors are then compared to templates of the candidate patterns, e.g., sounds or images to be identified in the signal.
  • the candidate patterns can represent e.g., names in a phonebook.
  • pattern recognition such as speech recognition is computationally demanding. In many cases, for example when implemented in embedded devices, due to the limited amount of memory and computational power there is a need to reduce the complexity of the pattern recognition algorithm.
  • the computational complexity depends on several factors: the sampling rate, the number of candidate model templates, and the feature vector dimension. Reducing any of these results in faster recognition that can be run in reasonable time on a certain processor, but this can result in poorer recognition accuracy.
  • Feature vector down-sampling A technique that reduces the decoding complexity by using the state likelihood (SL) measure corresponding to an incoming feature vector in several consecutive frames (time instants) .
  • Clustering of the model templates This technique clusters the acoustic space off-line. During decoding, a quick search among the clusters is performed first, and then only the SL measures for the members of the best matching cluster are evaluated.
  • An objective of the present invention is to solve or mitigate the above problems.
  • this objective is achieved by a method of the kind mentioned by way of introduction, further comprising formulating a control signal based on at least one time-dependent variable of the recognition process, and, for said at least one feature vector, computing only a subset of said distortion measure contributions using the vector components of said at least one feature vector, said subset being chosen in accordance with said control signal .
  • the expression "only a subset" indicates a number less than the number of distortion measure contributions available.
  • the subset includes less contributions than are defined in the comparison of the feature vector and the templates. This reduces the computational complexity of the computation, as the dimensionality of the vectors involved in the computation is effectively reduced. Although such a dimension reduction decreases the computational need, it has been found not to significantly impair the performance or noise robustness of the speech recognizer.
  • the solution according to the invention can reduce the complexity of the calculations by reducing the number of operations in the computation of the state likelihood, e.g. b-probability, that is a dominant factor in the computation process.
  • the solution according to the invention does not need extensive amounts of memory.
  • an embodiment of the invention may even operate without any additional memory depending on the actual implementation.
  • control signal is indicative of the processor load.
  • the reduction of complexity is thus adjusted according to the instantaneous processor capacity. This is a potential advantage of the present invention.
  • the control signal can alternatively be indicative of incoming signal properties.
  • the inventive concept can be viewed as masking some components of the feature vector itself, as it is the feature vectors that contain the information to be recognized.
  • the method can comprise a masking process, where some components of each feature vector are masked, by applying a mask.
  • the mask can omit selected components of the vectors, resulting in a reduced number of computed distortion measure contributions.
  • the component mask can be selected from a set of predefined masks, including at least one non-null mask, in accordance with said control signal. This results in an implementation requiring very little additional processing capacity to handle the masks.
  • the mask is dynamically computed in accordance with the control signal in each specific instance, resulting in a slightly more memory efficient implementation. Also, this implementation is more flexible, as the masks can be computed to match changing processing needs.
  • a set of masks is available, either by being stored in the memory of a pattern recognition device or by creating the masks dynamically, when necessary.
  • a mask from this set of masks is used to reduce the number of computed distortion measure contributions.
  • the speech recognition process can adapt to e.g., varying processor capacity, while still maintaining good recognition accuracy in low-load situations (and instants) .
  • Switching between masks can be performed even at a very high temporal resolution (e.g. frame-by-frame, every 10ms) . Therefore, it provides the maximum performance when the CPU is idle, and gives a graceful degradation when other load is present.
  • the mask may, at given intervals, mask all components in the feature vector, i.e., eliminating the entire distortion measure relating to this feature vector, and thereby causing a decimation of the sequence of feature vectors.
  • This offers the possibility to combine selective reduction of vector dimension with time-domain complexity reduction techniques, such as feature vector down-sampling.
  • specific vector components of successive feature vectors are used with a rate depending on their temporal characteristics. This makes it possible to achieve a feature component specific down-sampling, where feature components that, e.g., vary slowly in time can be down-sampled more than feature components varying rapidly in time.
  • Such down-sampling schemes can be implemented by properly adjusting the process of calculating and/or dynamically selecting the mask.
  • the subset of distortion measure contributions is combined with contributions from a previously computed distortion measure.
  • contributions from masked components that were skipped in the computation are replaced by the contributions from the most recently performed calculation of corresponding components.
  • this technique ensures that all distortion measures are calculated based on vectors of the same dimension. This simplifies future processing, e.g., eliminates the need of scaling when comparing distortion measures and the need of recalculating any constants dependent upon the number of contributions.
  • the invention can preferably be implemented in a speech recognition process, in which case the signal represents speech and the pattern represents spoken words.
  • the invention can advantageously be used in speech recognition systems implemented in embedded devices, such as mobile phones.
  • the templates can be Gaussian mixture densities of hidden Markov models (HMM) .
  • HMM hidden Markov models
  • the above objective is achieved with a device for pattern recognition, comprising means for forming a sequence of feature vectors from a digitized incoming signal, means for formulating a control signal based on at least one time-dependent variable of the recognition process, and means for comparing at least one feature vector with templates of candidate patterns by computing a distortion measure comprising distortion measure contributions wherein the comparing means are arranged to compute only a subset of the distortion measure contributions, the subset being chosen in accordance with said control signal .
  • Fig 1 illustrates the block diagram of a speech recognition engine
  • Fig 2a illustrates schematically computation of a distortion measure according to prior art.
  • Fig 2b illustrates schematically computation of a distortion measure according to an embodiment of the invention.
  • Fig 2c illustrates different masks suitable for the computation in fig 2b.
  • Fig 2d illustrates schematically computation of a distortion measure according to a second embodiment of the invention.
  • Fig 3 is a schematic flow chart of the masking process according to an embodiment of the invention.
  • Fig 4 illustrates a masking process according to a further embodiment of the invention.
  • Fig 5 illustrates a masking process according to a yet another embodiment of the invention.
  • the pattern recognizing process is a speech recognition process, used in e.g. voice based user interfaces.
  • voice based user interfaces e.g. voice based user interfaces
  • the incoming signals may be any digitized signals
  • the candidate patterns may represent sounds, images, texts, handwritten characters, etc.
  • a speech recognizer 1 as illustrated in fig 1 typically comprises a front-end processing section 2, responsible for the feature extraction, and a back-end processing section 3, responsible for the statistical analysis of extracted features with respect to model templates of candidate words or parts of words. These models can be created by on-line training (speaker- dependent name dialing, SDND) or by off-line training (speaker-independent name dialing, SIND) .
  • the input to a speech recognizer 1 consists of a digitally sampled waveform 4 split into consecutive, possibly overlapping segments. For each segment three main processing steps are performed:
  • Viterbi "decoding”, i.e., the current best cumulative distortion values 8 are obtained based on the distortion table computed in step S2 and the best cumulative distortion values for the previous speech segment 10.
  • the allowed transitions are constrained by the recognition lexicon plus grammar 9.
  • the current best recognition hypothesis as found by the Viterbi decoding step, is typically presented to the user as the recognition result.
  • Each acoustic model is usually represented by a hidden Markov model (HMM) .
  • the HMMs are the building blocks for the possible classification outcomes.
  • the HMM is a statistical automaton, which can accept/generate feature vectors. It consists of a set of states, and a set of allowed transitions between these states. Each transition has an associated probability value. Each state is described by a probability density function (PDF) on the space of feature vectors. The negative log-likelihood given by the state PDF and the feature vector can be also viewed as a distortion measure . Given the current state of the automaton it accepts/generates the current feature vector according to the likelihood given by the current state's PDF and then makes a transition to a new state as constrained by the set of transition probabilities.
  • PDF probability density function
  • the HMM that, during time, results in the smallest aggregate distortion is selected as the recognition result .
  • One of the most demanding computations consists of evaluating, for every feature vector, the distortion to the states of the recognition models. As mentioned before, this distortion is normally computed as a state likelihood measure, (its value also referred to as "b- probability" ) .
  • the PDF of each state is a mixture of a certain number of Gaussian densities (e.g., 8) . Each density contains a mean and an inverse standard deviation parameter vector.
  • E is the log-likelihood of the density
  • x is the i th vector component of the feature vector
  • istd 2 denote the i mean and inverse standard deviation vector component
  • D represents the number of feature components (the feature vector dimension)
  • C is an additive constant equal to the logarithm of the product of inverse standard deviations times l/sqrt(2*pi) to the power of D, where D is the feature vector dimension.
  • W ⁇ and L are, respectively, the log-mixture weight and the log-likelihood for density i, M stands for the number of densities in the state and b is the b- probability value.
  • the results are stored in a so called b- probability table, needed by the Viterbi algorithm.
  • This algorithm is used to determine a sequence of HMMs which best matches, in the maximum likelihood sense, the stream of input feature vectors.
  • the algorithm is implemented using a dynamic programming methodology. The number of multiplications and additions required to compute the b-probability table can be approximated as follows :
  • the number of required operations is reduced by masking some of the vector components, so that they are not taken into account in eq. 1.
  • the complexity reduction will be approximately proportional to the relative number of masked components. This is illustrated in fig 2b. In this case, a mask 21 is allowed to reduce the number of computed distortion measure contributions 23.
  • the black sections of the mask 21 indicate terms 23 in eq. 1 that are not computed. As a result, some terms 24 of the set 20 are masked, and only the remaining terms 22 are computed and summed to generate the log-likelihood value L.
  • the vectors ( ⁇ , istd, x) and sets (20) in figs 2a and 2b are schematic, and that each marked section in reality can comprise several components or terms.
  • the masking as described above relates to the log-likelihood terms 23, the masking process can also be viewed as if the feature vector, x, and the density vectors, ⁇ and istd, were reduced, this in turn leading to a reduced number of distortion measure contributions.
  • the masks may vary with time, i.e., different contributions are to be masked in different frames.
  • the variations can be based on, e.g., current processor load or properties of the input signal.
  • the recognizer 1 in fig 1 can be provided with (or connected to) a detector 11 of such load or properties, arranged to generate a control signal 12 used in the distortion computation S2.
  • Fig 2c shows three different masks 21 that can be applied to the computation in fig 2b.
  • Each mask 21 has a different scope, i.e., able to reduce the number of distortion measure contributions 23 by a different factor, x, y and z respectively.
  • the masked terms 24 can then be replaced by corresponding, previously computed terms 25, as illustrated in fig 2d.
  • a flow chart of the masking process implemented in a recognition process as described above, is illustrated in fig 3.
  • the flow chart relates to the handling of one feature vector.
  • a control signal is formulated.
  • this control signal can be indicative of the processor load, or any other time dependent variable of the recognizing process.
  • an appropriate mask 21 is selected.
  • the feature masks can be pre-computed for certain pre-defined complexity levels
  • the recognizer 1 is provided with software adapted to calculate a suitable mask for each instance, or at least at regular intervals. Such software can be implemented in the front-end 2, and/or in the back-end 3.
  • An advantage with computing the masks dynamically is that the masking process then is more adaptive to varying conditions.
  • the number and scope of the optimal masking factors may change depending on application, type of recognition, environment, etc.
  • step Sll can make the mask selection for each feature vector individually, or take several feat.ure vectors into account in a more elaborate feature vector reduction. Examples of such more elaborate schemes are given in the following description.
  • step S12 a set of density vectors is loaded from the static memory for comparison with the current feature vector.
  • the constant C in Eq. 1 is dependent on the selection of computed log-likelihood terms 23. Therefore, for each particular feature mask a corresponding C value is required for the loaded densities. Such a C value is determined in step S13.
  • step S13 is a simple selection from these values.
  • the relative memory increase resulting from this storage can be approximated as :
  • N is the number of masks
  • dim is the number of feature vector components.
  • the C values are not stored, in order to save memory. Instead, they can be re- computed in step S13 every time a density is loaded for processing from the static memory. In such a scenario, feature vector masking can be implemented without any need for extra memory at all.
  • step S14 the log-likelihood L according to eq.l is calculated.
  • the masking process has the effect to skip some of the terms in the summation in eq. 1, thereby reducing the calculation complexity, as shown in fig 2b.
  • the mask is simply a set of rules defining which terms 23 to skip and which terms 22 to compute during the calculation of eq.l.
  • the step S14 may also include the completion of the distortion measure contributions as was described above with reference to fig 2d. This can eliminate the need for step S13, as a full scale summation is performed in this case.
  • step S15 directs program control back to step S12 and loads the next set of density vectors in the state. This is repeated for all densities in the state.
  • step S16 the b-probability (eq. 2) can be calculated and stored in the b-probability table, and step S17 then directs program control back to step S12 and loads the first set of density vectors in the next state. This is repeated for all states.
  • step S18 the Viterbi algorithm is implemented in a manner known per se .
  • the masking can be adapted to include the principle of feature down- sampling, by "masking" the entire vector of selected frames. While feature down-sampling removes time-domain redundancies by decimating the features (e.g., by a factor of 2), feature component masking according to the above description eliminates the least useful feature vector components in every feature sample.
  • the feature vectors may be formed by concatenating components extracted from various sources. Due to this the feature vector space is in fact a product of the sub- spaces of the individual sources. For the majority of cases the distortion measure can be factored into several terms by taking advantage of the decomposition of the feature space. Since for the classification algorithm a sequence of feature vectors needs to be processed, in another embodiment of the invention the feature space is first partitioned into two subspaces; one with rapidly varying and another with slowly varying components. Since for each density the distortion is obtained by combining, possibly by weighting, the distortion of the two sub- spaces, the method can effectively reduce the computation by down-sampling the slowly varying components.
  • each feature vector 41-47 can be divided into 3 parts a, b, c: a very slowly varying subspace, a, for down-sampling by 2, a slowly varying subspace, b, for down-sampling by 3/2, and a rapidly varying subspace, c, for no down- sampling.
  • This is achieved by masking different components with different periodicity.
  • the b subspace is masked in every third frame, and the a subspace is masked in every second frame.
  • the decomposition is done prior to the recognition process, e.g., by analyzing the feature stream.
  • the components are assigned to the appropriate sub-spaces.
  • the degree of variation can also be known a priori from the front-end design (e.g. multi-resolution front-end, or a front-end combining features of different types) .
  • the decision to compute or not a given distortion measure for a certain sub-space can also be controlled by using a similarity measure between successive features.
  • the sub-space decomposition is done entirely at run time, for pairs or groups of features, with the help of a similarity measure. For every feature component, a one-dimensional similarity measure is computed. The components with the slowest variation (as indicated by the similarity measure) are placed in the slow varying subspace for the given group of features .
  • Figure 6 describes how the pre-defined or dynamically computed feature masks can be selected according to the actual current load of the processor in the recognizer 1.
  • Thresholds Thl, Th2 and Th3 are predefined, and when they are reached, the complexity of the probability computation is altered by switching between feature masks 1, 2, or 3 , having different masking factors.
  • feature masking can be completely disabled (dis) to provide maximum accuracy. Note that figure 6 does not include the load caused by the recognizer engine itself.
  • the process of mask selection can also be adaptive, so that the actual load resulting from a certain mask selection is used to determine what mask to use. By employing a learning algorithm, the impact of different masks on the CPU load can be registered, in order to improve future mask selections.
  • the selection of feature vector masks is performed based on input signal properties, e.g., time variation, signal/noise-ratio, etc. The input signal is first analyzed, and a suitable mask is selected in accordance with the determined properties.
  • the result of the recognition process is analyzed, and the mask selection is performed and adapted based on these results. For a particular mask, it is determined if the recognition results are satisfactory. If not, a less extensive mask is selected. If the masking is satisfactory, the mask is maintained, or even exchanged for a more extensive mask. In other words, the success of recognition is maintained at a desired level, while masking as many vector components as possible.

Abstract

Pattern recognition, wherein a sequence of feature vectors is formed from a digitized incoming signal, the feature vectors comprising feature vector components, and at least one feature vector is compared with templates of candidate patterns by computing a distortion measure. According to the invention, a control signal based on at least one time-dependent variable of the recognition process is formulated, and the distortion measure is computed using only a subset of the vector components of the feature vector, the subset being chosen in accordance with said control signal. This reduces the computational complexity of the computation , as the dimensionality of the vectors involved in the computation is effectively reduced. Although such a dimension reduction decreases the computational need, it has found not to significantly impair the classification performance.

Description

PATTERN RECOGNITION
Technical Field
The present invention relates to pattern recognition, where a set of feature vectors is formed from digitized incoming signals, and compared with templates of candidate patterns.
Technical Background
In pattern recognition, incoming signals are digitized, and a sequence of feature vectors are formed. These feature vectors are then compared to templates of the candidate patterns, e.g., sounds or images to be identified in the signal. In the case of speech recognition, the candidate patterns can represent e.g., names in a phonebook. However, pattern recognition such as speech recognition is computationally demanding. In many cases, for example when implemented in embedded devices, due to the limited amount of memory and computational power there is a need to reduce the complexity of the pattern recognition algorithm.
The computational complexity depends on several factors: the sampling rate, the number of candidate model templates, and the feature vector dimension. Reducing any of these results in faster recognition that can be run in reasonable time on a certain processor, but this can result in poorer recognition accuracy.
Furthermore, available resources are usually shared between different processes, and the available processing power and memory capacity is therefore variable. If the recognition functionality of an embedded device, having limited processing capacity to begin with, is to work at all times, it is even more crucial to minimize or dynamically adjust the processing requirements, without losing recognition accuracy. Conventional complexity reduction of pattern recognizers has been addressed by at least the following prior art techniques :
1. Feature vector down-sampling A technique that reduces the decoding complexity by using the state likelihood (SL) measure corresponding to an incoming feature vector in several consecutive frames (time instants) .
2. Clustering of the model templates This technique clusters the acoustic space off-line. During decoding, a quick search among the clusters is performed first, and then only the SL measures for the members of the best matching cluster are evaluated.
3. Lowering the feature vector dimension The number of feature vector components are reduced to a predefined number, using advanced linear transforms, such as PCA, LDA, etc, or neural networks.
Focusing on the third category, conventional examples of this technique do not have the flexibility to scale the computational complexity according to the available CPU power. Instead, it is always considered with the worst-case scenario. In addition, spectro- temporal linear transforms or neural network-based mappings may significantly increase the complexity of the front-end, and thus the whole recognizer.
An example of feature vector dimension reduction is given in "Should recognizers have ears", Speech Communication, Vol.25, pp. 3-27, 1998.
Summary Disclosure of the Invention
An objective of the present invention is to solve or mitigate the above problems.
According to a first aspect of the invention, this objective is achieved by a method of the kind mentioned by way of introduction, further comprising formulating a control signal based on at least one time-dependent variable of the recognition process, and, for said at least one feature vector, computing only a subset of said distortion measure contributions using the vector components of said at least one feature vector, said subset being chosen in accordance with said control signal .
It should be emphasized that the expression "only a subset" indicates a number less than the number of distortion measure contributions available. In other words, the subset includes less contributions than are defined in the comparison of the feature vector and the templates. This reduces the computational complexity of the computation, as the dimensionality of the vectors involved in the computation is effectively reduced. Although such a dimension reduction decreases the computational need, it has been found not to significantly impair the performance or noise robustness of the speech recognizer.
More specifically, the solution according to the invention can reduce the complexity of the calculations by reducing the number of operations in the computation of the state likelihood, e.g. b-probability, that is a dominant factor in the computation process.
Further, the solution according to the invention does not need extensive amounts of memory. In fact, an embodiment of the invention may even operate without any additional memory depending on the actual implementation.
According to one embodiment, the control signal is indicative of the processor load. The reduction of complexity is thus adjusted according to the instantaneous processor capacity. This is a potential advantage of the present invention. The control signal can alternatively be indicative of incoming signal properties.
The inventive concept can be viewed as masking some components of the feature vector itself, as it is the feature vectors that contain the information to be recognized. With this terminology, the method can comprise a masking process, where some components of each feature vector are masked, by applying a mask. The mask can omit selected components of the vectors, resulting in a reduced number of computed distortion measure contributions.
The component mask can be selected from a set of predefined masks, including at least one non-null mask, in accordance with said control signal. This results in an implementation requiring very little additional processing capacity to handle the masks.
Alternatively, the mask is dynamically computed in accordance with the control signal in each specific instance, resulting in a slightly more memory efficient implementation. Also, this implementation is more flexible, as the masks can be computed to match changing processing needs.
In other words, a set of masks is available, either by being stored in the memory of a pattern recognition device or by creating the masks dynamically, when necessary. Depending on the control signal, a mask from this set of masks is used to reduce the number of computed distortion measure contributions. As using different masks results in different computational complexity, the speech recognition process can adapt to e.g., varying processor capacity, while still maintaining good recognition accuracy in low-load situations (and instants) . Switching between masks can be performed even at a very high temporal resolution (e.g. frame-by-frame, every 10ms) . Therefore, it provides the maximum performance when the CPU is idle, and gives a graceful degradation when other load is present.
If deemed advantageous, the mask may, at given intervals, mask all components in the feature vector, i.e., eliminating the entire distortion measure relating to this feature vector, and thereby causing a decimation of the sequence of feature vectors. This offers the possibility to combine selective reduction of vector dimension with time-domain complexity reduction techniques, such as feature vector down-sampling. According to one embodiment, specific vector components of successive feature vectors are used with a rate depending on their temporal characteristics. This makes it possible to achieve a feature component specific down-sampling, where feature components that, e.g., vary slowly in time can be down-sampled more than feature components varying rapidly in time. Such down-sampling schemes can be implemented by properly adjusting the process of calculating and/or dynamically selecting the mask.
According to yet another embodiment, the subset of distortion measure contributions is combined with contributions from a previously computed distortion measure. In other words, contributions from masked components that were skipped in the computation, are replaced by the contributions from the most recently performed calculation of corresponding components. This means that a non-computed contribution is approximated with the most recently calculated, corresponding contribution, improving the performance without significantly increasing computational cost. Also, this technique ensures that all distortion measures are calculated based on vectors of the same dimension. This simplifies future processing, e.g., eliminates the need of scaling when comparing distortion measures and the need of recalculating any constants dependent upon the number of contributions. The invention can preferably be implemented in a speech recognition process, in which case the signal represents speech and the pattern represents spoken words. The invention can advantageously be used in speech recognition systems implemented in embedded devices, such as mobile phones. Further, the templates can be Gaussian mixture densities of hidden Markov models (HMM) . According to a second aspect of the invention, the above objective is achieved with a device for pattern recognition, comprising means for forming a sequence of feature vectors from a digitized incoming signal, means for formulating a control signal based on at least one time-dependent variable of the recognition process, and means for comparing at least one feature vector with templates of candidate patterns by computing a distortion measure comprising distortion measure contributions wherein the comparing means are arranged to compute only a subset of the distortion measure contributions, the subset being chosen in accordance with said control signal .
Brief Description of the Drawings
These and other aspects of the invention will be apparent from the preferred embodiments more clearly described with reference to the appended drawings.
Fig 1 illustrates the block diagram of a speech recognition engine
Fig 2a illustrates schematically computation of a distortion measure according to prior art.
Fig 2b illustrates schematically computation of a distortion measure according to an embodiment of the invention.
Fig 2c illustrates different masks suitable for the computation in fig 2b.
Fig 2d illustrates schematically computation of a distortion measure according to a second embodiment of the invention.
Fig 3 is a schematic flow chart of the masking process according to an embodiment of the invention.
Fig 4 illustrates a masking process according to a further embodiment of the invention. Fig 5 illustrates a masking process according to a yet another embodiment of the invention.
Fig 6 illustrates the effect of processor load. Detailed description of preferred embodiments In the following description, the pattern recognizing process is a speech recognition process, used in e.g. voice based user interfaces. However, this should not be regarded as a limitation to the invention, which is directed to pattern recognition in general. The incoming signals may be any digitized signals, and the candidate patterns may represent sounds, images, texts, handwritten characters, etc.
A speech recognizer 1 as illustrated in fig 1 typically comprises a front-end processing section 2, responsible for the feature extraction, and a back-end processing section 3, responsible for the statistical analysis of extracted features with respect to model templates of candidate words or parts of words. These models can be created by on-line training (speaker- dependent name dialing, SDND) or by off-line training (speaker-independent name dialing, SIND) . The input to a speech recognizer 1 consists of a digitally sampled waveform 4 split into consecutive, possibly overlapping segments. For each segment three main processing steps are performed:
51. Feature extraction, producing a vector of features 5.
52. Computation of the distortion values for the current feature vector compared to the acoustic model templates 6 (in the example below referred to as Guassian densities) , resulting in a distortion table 7 (in the example below referred to as a b-probability table) .
53. Viterbi "decoding", i.e., the current best cumulative distortion values 8 are obtained based on the distortion table computed in step S2 and the best cumulative distortion values for the previous speech segment 10. The allowed transitions are constrained by the recognition lexicon plus grammar 9. When the speech input ends, the current best recognition hypothesis, as found by the Viterbi decoding step, is typically presented to the user as the recognition result. Each acoustic model is usually represented by a hidden Markov model (HMM) . The HMMs are the building blocks for the possible classification outcomes.
The HMM is a statistical automaton, which can accept/generate feature vectors. It consists of a set of states, and a set of allowed transitions between these states. Each transition has an associated probability value. Each state is described by a probability density function (PDF) on the space of feature vectors. The negative log-likelihood given by the state PDF and the feature vector can be also viewed as a distortion measure . Given the current state of the automaton it accepts/generates the current feature vector according to the likelihood given by the current state's PDF and then makes a transition to a new state as constrained by the set of transition probabilities.
The HMM that, during time, results in the smallest aggregate distortion is selected as the recognition result .
One of the most demanding computations consists of evaluating, for every feature vector, the distortion to the states of the recognition models. As mentioned before, this distortion is normally computed as a state likelihood measure, (its value also referred to as "b- probability" ) . In a typical recognition engine, the PDF of each state is a mixture of a certain number of Gaussian densities (e.g., 8) . Each density contains a mean and an inverse standard deviation parameter vector.
During recognition, every incoming feature vector is first matched against the parameters (mean and standard deviation) of each density, to generate a distortion measure based on the log-likelihood value as follows, L = C - ∑(x, -μ, J - istd? , (1)
where E is the log-likelihood of the density, x is the ith vector component of the feature vector, μ and istd2 denote the i mean and inverse standard deviation vector component,
D represents the number of feature components (the feature vector dimension) , and
C is an additive constant equal to the logarithm of the product of inverse standard deviations times l/sqrt(2*pi) to the power of D, where D is the feature vector dimension.
The state b-probability is then given as follows
" 15 (2) b = log| p(^+Z,)
/=1 where Wλ and L are, respectively, the log-mixture weight and the log-likelihood for density i, M stands for the number of densities in the state and b is the b- probability value.
After calculating the b-probability values for all the states, the results are stored in a so called b- probability table, needed by the Viterbi algorithm. This algorithm is used to determine a sequence of HMMs which best matches, in the maximum likelihood sense, the stream of input feature vectors. The algorithm is implemented using a dynamic programming methodology. The number of multiplications and additions required to compute the b-probability table can be approximated as follows :
# multiplications = #all densities * #feature components *2, # additions = #multiplications .
The computation of the log-likelihood is illustrated in fig 2a. A density 6, comprising two vectors (μ and istd) of dimension N, and a feature vector 5, comprising one vector (x) of dimension N, are multiplied according to eq.l to form a set 20 of N log-likelihood terms 23, which are then summed according to eq.l to generate the value L. According to the invention, the number of required operations is reduced by masking some of the vector components, so that they are not taken into account in eq. 1. The complexity reduction will be approximately proportional to the relative number of masked components. This is illustrated in fig 2b. In this case, a mask 21 is allowed to reduce the number of computed distortion measure contributions 23. The black sections of the mask 21 indicate terms 23 in eq. 1 that are not computed. As a result, some terms 24 of the set 20 are masked, and only the remaining terms 22 are computed and summed to generate the log-likelihood value L. Note that the vectors (μ, istd, x) and sets (20) in figs 2a and 2b are schematic, and that each marked section in reality can comprise several components or terms. Although the masking as described above relates to the log-likelihood terms 23, the masking process can also be viewed as if the feature vector, x, and the density vectors, μ and istd, were reduced, this in turn leading to a reduced number of distortion measure contributions. Some advantages of the invention are more clearly understood when viewed this way, and some examples below, for example figures 4 and 5, actually refer to masking of the feature vectors x.
According to the invention, the masks may vary with time, i.e., different contributions are to be masked in different frames. The variations can be based on, e.g., current processor load or properties of the input signal. For this purpose, the recognizer 1 in fig 1 can be provided with (or connected to) a detector 11 of such load or properties, arranged to generate a control signal 12 used in the distortion computation S2. Fig 2c shows three different masks 21 that can be applied to the computation in fig 2b. Each mask 21 has a different scope, i.e., able to reduce the number of distortion measure contributions 23 by a different factor, x, y and z respectively.
In some cases it may be advantageous to preserve the number of components or terms in the log-likelihood computation (eq. 1) . The masked terms 24 can then be replaced by corresponding, previously computed terms 25, as illustrated in fig 2d.
A flow chart of the masking process implemented in a recognition process as described above, is illustrated in fig 3. The flow chart relates to the handling of one feature vector. First, in step S10, a control signal is formulated. As mentioned, this control signal can be indicative of the processor load, or any other time dependent variable of the recognizing process. Then, in step Sll, an appropriate mask 21 is selected. The feature masks can be pre-computed for certain pre-defined complexity levels
(e.g., masking factors x, y, z) and they can even be hard coded into the front -end 3 software to minimize the computational overhead. Alternatively, the recognizer 1 is provided with software adapted to calculate a suitable mask for each instance, or at least at regular intervals. Such software can be implemented in the front-end 2, and/or in the back-end 3.
An advantage with computing the masks dynamically is that the masking process then is more adaptive to varying conditions. The number and scope of the optimal masking factors may change depending on application, type of recognition, environment, etc.
Note that step Sll can make the mask selection for each feature vector individually, or take several feat.ure vectors into account in a more elaborate feature vector reduction. Examples of such more elaborate schemes are given in the following description. In step S12, a set of density vectors is loaded from the static memory for comparison with the current feature vector.
As described above, the constant C in Eq. 1 is dependent on the selection of computed log-likelihood terms 23. Therefore, for each particular feature mask a corresponding C value is required for the loaded densities. Such a C value is determined in step S13.
The required C constants can be pre-computed and stored in the memory of the recognizer, so that step S13 is a simple selection from these values. The relative memory increase resulting from this storage can be approximated as :
delta = dim+1
where delta is the required memory increase, N is the number of masks, and dim is the number of feature vector components. In a practical scenario with three masks, e.g., for complexities of 80%, 60%, and 40%, and with a feature vector dimension of 39, the memory increase would be 3/ (39+1) =7.5% of the size of the acoustic models. Alternatively, the C values are not stored, in order to save memory. Instead, they can be re- computed in step S13 every time a density is loaded for processing from the static memory. In such a scenario, feature vector masking can be implemented without any need for extra memory at all.
In step S14, the log-likelihood L according to eq.l is calculated. The masking process has the effect to skip some of the terms in the summation in eq. 1, thereby reducing the calculation complexity, as shown in fig 2b. The mask is simply a set of rules defining which terms 23 to skip and which terms 22 to compute during the calculation of eq.l. The step S14 may also include the completion of the distortion measure contributions as was described above with reference to fig 2d. This can eliminate the need for step S13, as a full scale summation is performed in this case.
After the log-likelihood E has been computed in step S14, step S15 directs program control back to step S12 and loads the next set of density vectors in the state. This is repeated for all densities in the state. After all densities in a state have been compared to the current feature vector, in step S16 the b-probability (eq. 2) can be calculated and stored in the b-probability table, and step S17 then directs program control back to step S12 and loads the first set of density vectors in the next state. This is repeated for all states.
Finally, in step S18, the Viterbi algorithm is implemented in a manner known per se .
According to one embodiment, the masking can be adapted to include the principle of feature down- sampling, by "masking" the entire vector of selected frames. While feature down-sampling removes time-domain redundancies by decimating the features (e.g., by a factor of 2), feature component masking according to the above description eliminates the least useful feature vector components in every feature sample.
Even for a given masking percentage, it is possible to select the most suitable components to mask at a given moment in time, and these components may vary with time. An example is given in fig 4, where five feature vectors, 31-35, are masked in an alternating manner, resulting in a constant masking factor, but with no component being masked more than every fourth frame .
Further, in many pattern recognition applications the feature vectors may be formed by concatenating components extracted from various sources. Due to this the feature vector space is in fact a product of the sub- spaces of the individual sources. For the majority of cases the distortion measure can be factored into several terms by taking advantage of the decomposition of the feature space. Since for the classification algorithm a sequence of feature vectors needs to be processed, in another embodiment of the invention the feature space is first partitioned into two subspaces; one with rapidly varying and another with slowly varying components. Since for each density the distortion is obtained by combining, possibly by weighting, the distortion of the two sub- spaces, the method can effectively reduce the computation by down-sampling the slowly varying components. This corresponds to selecting slightly different masks for different frames, possibly with different masking percentages . The number of sub-spaces is not necessarily two. A more refined decomposition can be achieved if the degree of variation is large. As illustrated in fig 5, each feature vector 41-47 can be divided into 3 parts a, b, c: a very slowly varying subspace, a, for down-sampling by 2, a slowly varying subspace, b, for down-sampling by 3/2, and a rapidly varying subspace, c, for no down- sampling. This is achieved by masking different components with different periodicity. The b subspace is masked in every third frame, and the a subspace is masked in every second frame. As shown in fig 5, this results in three different masks 48-50, applied in combination with a null mask 51 for some frames.
For the selection of the sub-space decomposition and down-sampling rates there are the following alternatives: 1. Static selection
2. Semi-dynamic selection
3. Dynamic selection
In the static selection process, the decomposition is done prior to the recognition process, e.g., by analyzing the feature stream. In the simplest approach, based on the cepstral information for each component the components are assigned to the appropriate sub-spaces. The degree of variation can also be known a priori from the front-end design (e.g. multi-resolution front-end, or a front-end combining features of different types) .
In the semi-dynamic selection, in addition to the static sub-space separation, the decision to compute or not a given distortion measure for a certain sub-space can also be controlled by using a similarity measure between successive features.
In the dynamic case the sub-space decomposition is done entirely at run time, for pairs or groups of features, with the help of a similarity measure. For every feature component, a one-dimensional similarity measure is computed. The components with the slowest variation (as indicated by the similarity measure) are placed in the slow varying subspace for the given group of features .
Figure 6 describes how the pre-defined or dynamically computed feature masks can be selected according to the actual current load of the processor in the recognizer 1. Thresholds Thl, Th2 and Th3 are predefined, and when they are reached, the complexity of the probability computation is altered by switching between feature masks 1, 2, or 3 , having different masking factors. When not needed, i.e., when processor load does not exceed threshold 1, feature masking can be completely disabled (dis) to provide maximum accuracy. Note that figure 6 does not include the load caused by the recognizer engine itself.
The process of mask selection can also be adaptive, so that the actual load resulting from a certain mask selection is used to determine what mask to use. By employing a learning algorithm, the impact of different masks on the CPU load can be registered, in order to improve future mask selections. According to another embodiment of the invention, the selection of feature vector masks is performed based on input signal properties, e.g., time variation, signal/noise-ratio, etc. The input signal is first analyzed, and a suitable mask is selected in accordance with the determined properties.
Alternatively, the result of the recognition process is analyzed, and the mask selection is performed and adapted based on these results. For a particular mask, it is determined if the recognition results are satisfactory. If not, a less extensive mask is selected. If the masking is satisfactory, the mask is maintained, or even exchanged for a more extensive mask. In other words, the success of recognition is maintained at a desired level, while masking as many vector components as possible.
As apparent to the skilled person, a number of modifications and variations of the above described embodiments are possible within the scope of the appended claims. For example, other types of control signals may be employed, in order to optimize the masking process. Other types of criteria for selecting masks and the scope of the masks can also be envisaged. Also, the described computation of distortion measures can be employed in pattern recognition processes different from the one described herein. Finally, the described methods can equally well be applied with other types of distortion measures as long as they can be formed by using partial distortion values from the component subspaces .

Claims

1. A method for pattern recognition, wherein a sequence of feature vectors (5) is formed from a digitized incoming signal (4), said feature vectors (5) comprising feature vector components, and at least one feature vector is compared with templates (6) of candidate patterns by computing a distortion measure (L) including distortion measure contributions (23) , said method being characteri zed by formulating (S10) a control signal (12) based on at least one time-dependent variable of the recognition process, and for said at least one feature vector, computing (S14) only a subset (22) of said distortion measure (L) contributions (23) using the vector components of said at least one feature vector, said subset (22) being chosen in accordance with said control signal (12) .
2. A method according to claim 1, wherein said time-dependent variable belongs to the group of processor load and incoming signal (4) properties.
3. A method according to claim 1 or 2 , wherein the vector components of said at least one feature vector are masked, by applying a mask (21) , thereby reducing the number of computed distortion measure contributions in the subset (22) .
4. A method according to claim 3, wherein said mask is dynamically computed (Sll) in accordance with said control signal (12) .
5. A method according to claim 3, wherein said mask (21) is selected (Sll) from a set of predefined masks in accordance with said control signal (12) .
6. A method according to claim 5, wherein said predefined masks represent different reductions of the number of computed distortion measure contributions in the subset (22) .
7. A method according to claim 3 - 6, wherein said mask at given time instances comprises all vector components in the feature vector, causing a decimation of the sequence of feature vectors .
8. A method according to any one of the preceding claims, wherein specific vector components of successive feature vectors are used with a rate depending on their temporal characteristics.
9. A method according to any one of the preceding claims, wherein said subset (22) of distortion measure
(L) contributions (23) is combined with contributions (25) from a previously computed distortion measure.
10. A method according to any one of the preceding claims, wherein said signal (4) represents speech and the candidate patterns represent spoken utterances.
11. A method according to any one of the preceding claims, wherein said templates (6) are Gaussian mixture densities of Hidden Markov Models (HMMs) .
12. A method according to claim 11, wherein said distortion measure is based on a log-likelihood (E) .
13. A device (1) for pattern recognition, comprising means (2) for forming a sequence of feature vectors (5) from a digitized incoming signal, characterized by means (11) for formulating a control signal (12) based on at least one time-dependent variable of the recognition process, and means (3) for comparing at least one feature vector with templates (6) of candidate patterns by computing a distortion measure (L) comprising distortion measure contributions (23), wherein said comparing means are arranged to compute only a subset (22) of the distortion measure contributions (23) , said subset being chosen in accordance with said control signal (12) .
14. A device according to claim 13 , wherein said means (11) for formulating a control signal (12) is arranged to detect a processor load of the device (1) .
15. A device according to claim 13 or 14, further including means (3) for applying a mask (21) to the components of said at least one feature vector, thereby reducing the number of distortion measure contributions in the subset (22) .
16. A device according to claim 13 - 15, further including means (3) for selecting said mask from a set of predefined masks in accordance with said control signal (12) .
17. A device according to claim 13 - 15, further including means (3) for dynamically computing said mask in accordance with said control signal (12) .
18. A device according to claims 13 - 17, implemented as an embedded processing device, comprising a front -end section (2) for forming said sequence of feature vectors, and a back-end section (3) for providing said set of distortion measures ( ).
19. A speech recognizer (1) including a device according to claim 13 - 18.
20. A communication device comprising a speech recognizer according to claim 19.
21. A system for pattern recognition comprising means (2) for forming a sequence of feature vectors (5) from a digitized incoming signal, characteri zed by means (11) for formulating a control signal (12) based on at least one time-dependent variable of the recognition process, and means (3) for comparing at least one feature vector with templates (6) of candidate patterns by computing a distortion measure (E) comprising distortion measure contributions (23) , wherein said comparing means are arranged to compute only a subset (22) of the distortion measure contributions (23), said subset being chosen in accordance with said control signal (12) .
22. A computer program product, directly loadable into the memory of a computer, comprising computer program code means for performing the steps of the method of claims 1-12 when executed on the computer.
23. A computer program product according to claim 22, stored on a computer readable medium.
PCT/IB2002/000954 2002-03-27 2002-03-27 Pattern recognition WO2003081574A1 (en)

Priority Applications (7)

Application Number Priority Date Filing Date Title
PCT/IB2002/000954 WO2003081574A1 (en) 2002-03-27 2002-03-27 Pattern recognition
JP2003579211A JP4295118B2 (en) 2002-03-27 2002-03-27 Pattern recognition
EP02722529A EP1488408A1 (en) 2002-03-27 2002-03-27 Pattern recognition
KR1020047015018A KR100760666B1 (en) 2002-03-27 2002-03-27 Pattern recognition
AU2002253416A AU2002253416A1 (en) 2002-03-27 2002-03-27 Pattern recognition
CNB028286480A CN1295672C (en) 2002-03-27 2002-03-27 Pattern recognition
US10/402,367 US7269556B2 (en) 2002-03-27 2003-03-26 Pattern recognition

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/IB2002/000954 WO2003081574A1 (en) 2002-03-27 2002-03-27 Pattern recognition

Publications (1)

Publication Number Publication Date
WO2003081574A1 true WO2003081574A1 (en) 2003-10-02

Family

ID=28053150

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/IB2002/000954 WO2003081574A1 (en) 2002-03-27 2002-03-27 Pattern recognition

Country Status (7)

Country Link
US (1) US7269556B2 (en)
EP (1) EP1488408A1 (en)
JP (1) JP4295118B2 (en)
KR (1) KR100760666B1 (en)
CN (1) CN1295672C (en)
AU (1) AU2002253416A1 (en)
WO (1) WO2003081574A1 (en)

Families Citing this family (141)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6834308B1 (en) 2000-02-17 2004-12-21 Audible Magic Corporation Method and apparatus for identifying media content presented on a media playing device
US8645137B2 (en) 2000-03-16 2014-02-04 Apple Inc. Fast, language-independent method for user authentication by voice
US7562012B1 (en) 2000-11-03 2009-07-14 Audible Magic Corporation Method and apparatus for creating a unique audio signature
WO2002082271A1 (en) * 2001-04-05 2002-10-17 Audible Magic Corporation Copyright detection and protection system and method
US7529659B2 (en) * 2005-09-28 2009-05-05 Audible Magic Corporation Method and apparatus for identifying an unknown work
US8972481B2 (en) 2001-07-20 2015-03-03 Audible Magic, Inc. Playlist generation method and apparatus
US7877438B2 (en) * 2001-07-20 2011-01-25 Audible Magic Corporation Method and apparatus for identifying new media content
WO2003085638A1 (en) * 2002-03-27 2003-10-16 Nokia Corporation Pattern recognition
US8332326B2 (en) * 2003-02-01 2012-12-11 Audible Magic Corporation Method and apparatus to identify a work received by a processing system
US8130746B2 (en) * 2004-07-28 2012-03-06 Audible Magic Corporation System for distributing decoy content in a peer to peer network
US20060142972A1 (en) * 2004-12-29 2006-06-29 Snap-On Incorporated System and method of using sensors to emulate human senses for diagnosing an assembly
JP4298672B2 (en) * 2005-04-11 2009-07-22 キヤノン株式会社 Method and apparatus for calculating output probability of state of mixed distribution HMM
US8677377B2 (en) 2005-09-08 2014-03-18 Apple Inc. Method and apparatus for building an intelligent automated assistant
US9318108B2 (en) 2010-01-18 2016-04-19 Apple Inc. Intelligent automated assistant
US8355913B2 (en) * 2006-11-03 2013-01-15 Nokia Corporation Speech recognition with adjustable timeout period
US8977255B2 (en) 2007-04-03 2015-03-10 Apple Inc. Method and system for operating a multi-function portable electronic device using voice-activation
US8006314B2 (en) 2007-07-27 2011-08-23 Audible Magic Corporation System for identifying content of digital data
US9330720B2 (en) 2008-01-03 2016-05-03 Apple Inc. Methods and apparatus for altering audio output signals
US8996376B2 (en) 2008-04-05 2015-03-31 Apple Inc. Intelligent text-to-speech conversion
US10496753B2 (en) 2010-01-18 2019-12-03 Apple Inc. Automatically adapting user interfaces for hands-free interaction
US20100030549A1 (en) 2008-07-31 2010-02-04 Lee Michael M Mobile device having human language translation capability with positional feedback
US9959870B2 (en) 2008-12-11 2018-05-01 Apple Inc. Speech recognition involving a mobile device
US8200489B1 (en) * 2009-01-29 2012-06-12 The United States Of America As Represented By The Secretary Of The Navy Multi-resolution hidden markov model using class specific features
US8788256B2 (en) 2009-02-17 2014-07-22 Sony Computer Entertainment Inc. Multiple language voice recognition
US8442833B2 (en) * 2009-02-17 2013-05-14 Sony Computer Entertainment Inc. Speech processing with source location estimation using signals from two or more microphones
US8442829B2 (en) * 2009-02-17 2013-05-14 Sony Computer Entertainment Inc. Automatic computation streaming partition for voice recognition on multiple processors with limited memory
US8199651B1 (en) 2009-03-16 2012-06-12 Audible Magic Corporation Method and system for modifying communication flows at a port level
US10241752B2 (en) 2011-09-30 2019-03-26 Apple Inc. Interface for a virtual digital assistant
US20120311585A1 (en) 2011-06-03 2012-12-06 Apple Inc. Organizing task items that represent tasks to perform
US9858925B2 (en) 2009-06-05 2018-01-02 Apple Inc. Using context information to facilitate processing of commands in a virtual assistant
US10241644B2 (en) 2011-06-03 2019-03-26 Apple Inc. Actionable reminder entries
US9431006B2 (en) 2009-07-02 2016-08-30 Apple Inc. Methods and apparatuses for automatic speech recognition
EP2306457B1 (en) * 2009-08-24 2016-10-12 Oticon A/S Automatic sound recognition based on binary time frequency units
US10276170B2 (en) 2010-01-18 2019-04-30 Apple Inc. Intelligent automated assistant
US10679605B2 (en) 2010-01-18 2020-06-09 Apple Inc. Hands-free list-reading by intelligent automated assistant
US10553209B2 (en) 2010-01-18 2020-02-04 Apple Inc. Systems and methods for hands-free notification summaries
US10705794B2 (en) 2010-01-18 2020-07-07 Apple Inc. Automatically adapting user interfaces for hands-free interaction
WO2011089450A2 (en) 2010-01-25 2011-07-28 Andrew Peter Nelson Jerram Apparatuses, methods and systems for a digital conversation management platform
US8682667B2 (en) 2010-02-25 2014-03-25 Apple Inc. User profiling for selecting user specific voice input processing information
US10762293B2 (en) 2010-12-22 2020-09-01 Apple Inc. Using parts-of-speech tagging and named entity recognition for spelling correction
US9262612B2 (en) 2011-03-21 2016-02-16 Apple Inc. Device access using voice authentication
US10672399B2 (en) 2011-06-03 2020-06-02 Apple Inc. Switching between text data and audio data based on a mapping
US10057736B2 (en) 2011-06-03 2018-08-21 Apple Inc. Active transport based notifications
US8994660B2 (en) 2011-08-29 2015-03-31 Apple Inc. Text correction processing
KR101300247B1 (en) * 2011-11-11 2013-08-26 경희대학교 산학협력단 Markov chain hidden conditional random fields model based pattern recognition method
US10134385B2 (en) 2012-03-02 2018-11-20 Apple Inc. Systems and methods for name pronunciation
US9483461B2 (en) 2012-03-06 2016-11-01 Apple Inc. Handling speech synthesis of content for multiple languages
US9280610B2 (en) 2012-05-14 2016-03-08 Apple Inc. Crowd sourcing information to fulfill user requests
US9721563B2 (en) 2012-06-08 2017-08-01 Apple Inc. Name recognition system
US9495129B2 (en) 2012-06-29 2016-11-15 Apple Inc. Device, method, and user interface for voice-activated navigation and browsing of a document
US9576574B2 (en) 2012-09-10 2017-02-21 Apple Inc. Context-sensitive handling of interruptions by intelligent digital assistant
US9547647B2 (en) 2012-09-19 2017-01-17 Apple Inc. Voice-based media searching
US9081778B2 (en) 2012-09-25 2015-07-14 Audible Magic Corporation Using digital fingerprints to associate data with a work
KR20230137475A (en) 2013-02-07 2023-10-04 애플 인크. Voice trigger for a digital assistant
US10652394B2 (en) 2013-03-14 2020-05-12 Apple Inc. System and method for processing voicemail
US9368114B2 (en) 2013-03-14 2016-06-14 Apple Inc. Context-sensitive handling of interruptions
WO2014144579A1 (en) 2013-03-15 2014-09-18 Apple Inc. System and method for updating an adaptive speech recognition model
US9922642B2 (en) 2013-03-15 2018-03-20 Apple Inc. Training an at least partial voice command system
US9582608B2 (en) 2013-06-07 2017-02-28 Apple Inc. Unified ranking with entropy-weighted information for phrase-based semantic auto-completion
WO2014197336A1 (en) 2013-06-07 2014-12-11 Apple Inc. System and method for detecting errors in interactions with a voice-based digital assistant
WO2014197334A2 (en) 2013-06-07 2014-12-11 Apple Inc. System and method for user-specified pronunciation of words for speech synthesis and recognition
WO2014197335A1 (en) 2013-06-08 2014-12-11 Apple Inc. Interpreting and acting upon commands that involve sharing information with remote devices
US10176167B2 (en) 2013-06-09 2019-01-08 Apple Inc. System and method for inferring user intent from speech inputs
EP3937002A1 (en) 2013-06-09 2022-01-12 Apple Inc. Device, method, and graphical user interface for enabling conversation persistence across two or more instances of a digital assistant
KR101809808B1 (en) 2013-06-13 2017-12-15 애플 인크. System and method for emergency calls initiated by voice command
JP6163266B2 (en) 2013-08-06 2017-07-12 アップル インコーポレイテッド Automatic activation of smart responses based on activation from remote devices
US9620105B2 (en) 2014-05-15 2017-04-11 Apple Inc. Analyzing audio input for efficient speech and music recognition
US10592095B2 (en) 2014-05-23 2020-03-17 Apple Inc. Instantaneous speaking of content on touch devices
US9502031B2 (en) 2014-05-27 2016-11-22 Apple Inc. Method for supporting dynamic grammars in WFST-based ASR
US9430463B2 (en) 2014-05-30 2016-08-30 Apple Inc. Exemplar-based natural language processing
US9633004B2 (en) 2014-05-30 2017-04-25 Apple Inc. Better resolution when referencing to concepts
US10078631B2 (en) 2014-05-30 2018-09-18 Apple Inc. Entropy-guided text prediction using combined word and character n-gram language models
US9734193B2 (en) 2014-05-30 2017-08-15 Apple Inc. Determining domain salience ranking from ambiguous words in natural speech
US9842101B2 (en) 2014-05-30 2017-12-12 Apple Inc. Predictive conversion of language input
US10170123B2 (en) 2014-05-30 2019-01-01 Apple Inc. Intelligent assistant for home automation
US9715875B2 (en) 2014-05-30 2017-07-25 Apple Inc. Reducing the need for manual start/end-pointing and trigger phrases
US9760559B2 (en) 2014-05-30 2017-09-12 Apple Inc. Predictive text input
US10289433B2 (en) 2014-05-30 2019-05-14 Apple Inc. Domain specific language for encoding assistant dialog
US9785630B2 (en) 2014-05-30 2017-10-10 Apple Inc. Text prediction using combined word N-gram and unigram language models
EP3480811A1 (en) 2014-05-30 2019-05-08 Apple Inc. Multi-command single utterance input method
US9338493B2 (en) 2014-06-30 2016-05-10 Apple Inc. Intelligent automated assistant for TV user interactions
US10659851B2 (en) 2014-06-30 2020-05-19 Apple Inc. Real-time digital assistant knowledge updates
US10446141B2 (en) 2014-08-28 2019-10-15 Apple Inc. Automatic speech recognition based on user feedback
US9818400B2 (en) 2014-09-11 2017-11-14 Apple Inc. Method and apparatus for discovering trending terms in speech requests
US10789041B2 (en) 2014-09-12 2020-09-29 Apple Inc. Dynamic thresholds for always listening speech trigger
US9668121B2 (en) 2014-09-30 2017-05-30 Apple Inc. Social reminders
US10127911B2 (en) 2014-09-30 2018-11-13 Apple Inc. Speaker identification and unsupervised speaker adaptation techniques
US9886432B2 (en) 2014-09-30 2018-02-06 Apple Inc. Parsimonious handling of word inflection via categorical stem + suffix N-gram language models
US10074360B2 (en) 2014-09-30 2018-09-11 Apple Inc. Providing an indication of the suitability of speech recognition
US9646609B2 (en) 2014-09-30 2017-05-09 Apple Inc. Caching apparatus for serving phonetic pronunciations
US10552013B2 (en) 2014-12-02 2020-02-04 Apple Inc. Data detection
US9711141B2 (en) 2014-12-09 2017-07-18 Apple Inc. Disambiguating heteronyms in speech synthesis
US9865280B2 (en) 2015-03-06 2018-01-09 Apple Inc. Structured dictation using intelligent automated assistants
US9721566B2 (en) 2015-03-08 2017-08-01 Apple Inc. Competing devices responding to voice triggers
US10567477B2 (en) 2015-03-08 2020-02-18 Apple Inc. Virtual assistant continuity
US9886953B2 (en) 2015-03-08 2018-02-06 Apple Inc. Virtual assistant activation
US9899019B2 (en) 2015-03-18 2018-02-20 Apple Inc. Systems and methods for structured stem and suffix language models
US9842105B2 (en) 2015-04-16 2017-12-12 Apple Inc. Parsimonious continuous-space phrase representations for natural language processing
US10083688B2 (en) 2015-05-27 2018-09-25 Apple Inc. Device voice control for selecting a displayed affordance
US10127220B2 (en) 2015-06-04 2018-11-13 Apple Inc. Language identification from short strings
US10101822B2 (en) 2015-06-05 2018-10-16 Apple Inc. Language input correction
US9578173B2 (en) 2015-06-05 2017-02-21 Apple Inc. Virtual assistant aided communication with 3rd party service in a communication session
US10255907B2 (en) 2015-06-07 2019-04-09 Apple Inc. Automatic accent detection using acoustic models
US10186254B2 (en) 2015-06-07 2019-01-22 Apple Inc. Context-based endpoint detection
US11025565B2 (en) 2015-06-07 2021-06-01 Apple Inc. Personalized prediction of responses for instant messaging
US10671428B2 (en) 2015-09-08 2020-06-02 Apple Inc. Distributed personal assistant
US10747498B2 (en) 2015-09-08 2020-08-18 Apple Inc. Zero latency digital assistant
US9697820B2 (en) 2015-09-24 2017-07-04 Apple Inc. Unit-selection text-to-speech synthesis using concatenation-sensitive neural networks
US11010550B2 (en) 2015-09-29 2021-05-18 Apple Inc. Unified language modeling framework for word prediction, auto-completion and auto-correction
US10366158B2 (en) 2015-09-29 2019-07-30 Apple Inc. Efficient word encoding for recurrent neural network language models
US11587559B2 (en) 2015-09-30 2023-02-21 Apple Inc. Intelligent device identification
US10691473B2 (en) 2015-11-06 2020-06-23 Apple Inc. Intelligent automated assistant in a messaging environment
US10049668B2 (en) 2015-12-02 2018-08-14 Apple Inc. Applying neural network language models to weighted finite state transducers for automatic speech recognition
US10223066B2 (en) 2015-12-23 2019-03-05 Apple Inc. Proactive assistance based on dialog communication between devices
US10446143B2 (en) 2016-03-14 2019-10-15 Apple Inc. Identification of voice inputs providing credentials
US9934775B2 (en) 2016-05-26 2018-04-03 Apple Inc. Unit-selection text-to-speech synthesis based on predicted concatenation parameters
US9972304B2 (en) 2016-06-03 2018-05-15 Apple Inc. Privacy preserving distributed evaluation framework for embedded personalized systems
US10249300B2 (en) 2016-06-06 2019-04-02 Apple Inc. Intelligent list reading
US10049663B2 (en) 2016-06-08 2018-08-14 Apple, Inc. Intelligent automated assistant for media exploration
DK179309B1 (en) 2016-06-09 2018-04-23 Apple Inc Intelligent automated assistant in a home environment
US10490187B2 (en) 2016-06-10 2019-11-26 Apple Inc. Digital assistant providing automated status report
US10509862B2 (en) 2016-06-10 2019-12-17 Apple Inc. Dynamic phrase expansion of language input
US10586535B2 (en) 2016-06-10 2020-03-10 Apple Inc. Intelligent digital assistant in a multi-tasking environment
US10192552B2 (en) 2016-06-10 2019-01-29 Apple Inc. Digital assistant providing whispered speech
US10067938B2 (en) 2016-06-10 2018-09-04 Apple Inc. Multilingual word prediction
DK201670540A1 (en) 2016-06-11 2018-01-08 Apple Inc Application integration with a digital assistant
DK179415B1 (en) 2016-06-11 2018-06-14 Apple Inc Intelligent device arbitration and control
DK179343B1 (en) 2016-06-11 2018-05-14 Apple Inc Intelligent task discovery
DK179049B1 (en) 2016-06-11 2017-09-18 Apple Inc Data driven natural language event detection and classification
US10043516B2 (en) 2016-09-23 2018-08-07 Apple Inc. Intelligent automated assistant
US11281993B2 (en) 2016-12-05 2022-03-22 Apple Inc. Model and ensemble compression for metric learning
US10593346B2 (en) 2016-12-22 2020-03-17 Apple Inc. Rank-reduced token representation for automatic speech recognition
DK201770383A1 (en) 2017-05-09 2018-12-14 Apple Inc. User interface for correcting recognition errors
DK201770439A1 (en) 2017-05-11 2018-12-13 Apple Inc. Offline personal assistant
DK179745B1 (en) 2017-05-12 2019-05-01 Apple Inc. SYNCHRONIZATION AND TASK DELEGATION OF A DIGITAL ASSISTANT
DK179496B1 (en) 2017-05-12 2019-01-15 Apple Inc. USER-SPECIFIC Acoustic Models
DK201770428A1 (en) 2017-05-12 2019-02-18 Apple Inc. Low-latency intelligent automated assistant
DK201770432A1 (en) 2017-05-15 2018-12-21 Apple Inc. Hierarchical belief states for digital assistants
DK201770431A1 (en) 2017-05-15 2018-12-20 Apple Inc. Optimizing dialogue policy decisions for digital assistants using implicit feedback
DK179560B1 (en) 2017-05-16 2019-02-18 Apple Inc. Far-field extension for digital assistant services
CN109308896B (en) * 2017-07-28 2022-04-15 江苏汇通金科数据股份有限公司 Voice processing method and device, storage medium and processor

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP0125422A1 (en) * 1983-04-13 1984-11-21 Texas Instruments Incorporated Speaker-independent word recognizer
US6009199A (en) * 1996-07-12 1999-12-28 Lucent Technologies Inc. Classification technique using random decision forests
US6405168B1 (en) * 1999-09-30 2002-06-11 Conexant Systems, Inc. Speaker dependent speech recognition training using simplified hidden markov modeling and robust end-point detection

Family Cites Families (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5274714A (en) * 1990-06-04 1993-12-28 Neuristics, Inc. Method and apparatus for determining and organizing feature vectors for neural network recognition
US5590242A (en) * 1994-03-24 1996-12-31 Lucent Technologies Inc. Signal bias removal for robust telephone speech recognition
US5751905A (en) * 1995-03-15 1998-05-12 International Business Machines Corporation Statistical acoustic processing method and apparatus for speech recognition using a toned phoneme system
US6009390A (en) 1997-09-11 1999-12-28 Lucent Technologies Inc. Technique for selective use of Gaussian kernels and mixture component weights of tied-mixture hidden Markov models for speech recognition
US6098040A (en) * 1997-11-07 2000-08-01 Nortel Networks Corporation Method and apparatus for providing an improved feature set in speech recognition by performing noise cancellation and background masking
US6178401B1 (en) 1998-08-28 2001-01-23 International Business Machines Corporation Method for reducing search complexity in a speech recognition system
KR100362853B1 (en) * 2000-05-17 2002-12-11 엘지전자 주식회사 device for decreasing vibration in compressor
KR100408524B1 (en) * 2001-08-22 2003-12-06 삼성전자주식회사 Speech recognition method and the apparatus thereof
WO2003085638A1 (en) * 2002-03-27 2003-10-16 Nokia Corporation Pattern recognition
JP4292837B2 (en) * 2002-07-16 2009-07-08 日本電気株式会社 Pattern feature extraction method and apparatus
KR100476103B1 (en) * 2002-08-09 2005-03-10 한국과학기술원 Implementation of Masking Algorithm Using the Feature Space Filtering

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP0125422A1 (en) * 1983-04-13 1984-11-21 Texas Instruments Incorporated Speaker-independent word recognizer
US6009199A (en) * 1996-07-12 1999-12-28 Lucent Technologies Inc. Classification technique using random decision forests
US6405168B1 (en) * 1999-09-30 2002-06-11 Conexant Systems, Inc. Speaker dependent speech recognition training using simplified hidden markov modeling and robust end-point detection

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
RAYMER M.L. ET AL: "Dimensionality reduction using genetic algorithms", IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, vol. 4, no. 2, July 2000 (2000-07-01), pages 164 - 171, XP002903027 *

Also Published As

Publication number Publication date
JP2005521106A (en) 2005-07-14
JP4295118B2 (en) 2009-07-15
AU2002253416A1 (en) 2003-10-08
KR100760666B1 (en) 2007-09-20
CN1623183A (en) 2005-06-01
US7269556B2 (en) 2007-09-11
US20040039572A1 (en) 2004-02-26
EP1488408A1 (en) 2004-12-22
CN1295672C (en) 2007-01-17
KR20040111428A (en) 2004-12-31

Similar Documents

Publication Publication Date Title
US7269556B2 (en) Pattern recognition
US5638486A (en) Method and system for continuous speech recognition using voting techniques
US5596679A (en) Method and system for identifying spoken sounds in continuous speech by comparing classifier outputs
US5822728A (en) Multistage word recognizer based on reliably detected phoneme similarity regions
EP0617827B1 (en) Composite expert
US6076053A (en) Methods and apparatus for discriminative training and adaptation of pronunciation networks
US5812973A (en) Method and system for recognizing a boundary between contiguous sounds for use with a speech recognition system
US5459815A (en) Speech recognition method using time-frequency masking mechanism
US5621848A (en) Method of partitioning a sequence of data frames
EP0706171A1 (en) Speech recognition method and apparatus
CA2275712A1 (en) Speech recognition system employing discriminatively trained models
US5734793A (en) System for recognizing spoken sounds from continuous speech and method of using same
WO1995034035A1 (en) Method of training neural networks used for speech recognition
US5794190A (en) Speech pattern recognition using pattern recognizers and classifiers
JP2006215564A (en) Method and apparatus for predicting word accuracy in automatic speech recognition systems
JP3298858B2 (en) Partition-based similarity method for low-complexity speech recognizers
Watanabe et al. High speed speech recognition using tree-structured probability density function
EP1576580A1 (en) Method of optimising the execution of a neural network in a speech recognition system through conditionally skipping a variable number of frames
AU2362495A (en) Speech-recognition system utilizing neural networks and method of using same
US7912715B2 (en) Determining distortion measures in a pattern recognition process
Matsui et al. Smoothed N-best-based speaker adaptation for speech recognition
Hu et al. A neural network based nonlinear feature transformation for speech recognition.
JP3105708B2 (en) Voice recognition device
PCSI et al. CMAyer* MJ Hunt f and DM Brookes it
El-Yazeed et al. Multi-codebook Vector Quantization Algorithm for Speaker Identification

Legal Events

Date Code Title Description
AK Designated states

Kind code of ref document: A1

Designated state(s): AE AG AL AM AT AU AZ BA BB BG BR BY BZ CA CH CN CO CR CU CZ DE DK DM DZ EC EE ES FI GB GD GE GH GM HR HU ID IL IN IS JP KE KG KP KR KZ LC LK LR LS LT LU LV MA MD MG MK MN MW MX MZ NO NZ OM PH PL PT RO RU SD SE SG SI SK SL TJ TM TN TR TT TZ UA UG US UZ VN YU ZA ZM ZW

AL Designated countries for regional patents

Kind code of ref document: A1

Designated state(s): GH GM KE LS MW MZ SD SL SZ TZ UG ZM ZW AM AZ BY KG KZ MD RU TJ TM AT BE CH CY DE DK ES FI FR GB GR IE IT LU MC NL PT SE TR BF BJ CF CG CI CM GA GN GQ GW ML MR NE SN TD TG

DFPE Request for preliminary examination filed prior to expiration of 19th month from priority date (pct application filed before 20040101)
121 Ep: the epo has been informed by wipo that ep was designated in this application
WWE Wipo information: entry into national phase

Ref document number: 2002722529

Country of ref document: EP

WWE Wipo information: entry into national phase

Ref document number: 1020047015018

Country of ref document: KR

WWE Wipo information: entry into national phase

Ref document number: 20028286480

Country of ref document: CN

Ref document number: 2003579211

Country of ref document: JP

WWP Wipo information: published in national office

Ref document number: 2002722529

Country of ref document: EP

WWP Wipo information: published in national office

Ref document number: 1020047015018

Country of ref document: KR