WO2018192186A1 - Speech recognition method and apparatus - Google Patents

Abstract

Disclosed are a speech recognition method and apparatus. The method comprises: acquiring a feature classification result of a speech signal to be recognized, wherein the feature classification result contains a pronunciation for describing a pronunciation feature of each speech signal frame and the probability of mapping each speech signal frame to a corresponding pronunciation (S201); if the probability of mapping any speech signal frame to a corresponding pronunciation satisfies a predetermined filtering rule, determining the pronunciation as a candidate pronunciation (S202); if the probability of mapping any of a predetermined number of speech signal frames adjacent to the speech signal frame to the candidate pronunciation satisfies the predetermined filtering rule, deleting the candidate pronunciation from pronunciations contained in the feature classification result (S203); if none of the probabilities of mapping the predetermined number of speech signal frames adjacent to the speech signal frame to the candidate pronunciation satisfies the predetermined filtering rule, retaining the candidate pronunciation in the pronunciations contained in the feature classification result (S204); and recognizing the speech signal based on a filtered feature classification result (S205). During the process of speech signal recognition, a recognition operation related to a pronunciation filtered out is not required to be executed, thereby improving the efficiency of speech recognition.

Description

Speech recognition method and device Technical field

The present invention relates to the field of computer technologies, and in particular, to a voice recognition method and apparatus.

Background technique

With the development of computer technology, the application of Automatic Speech Recognition (ASR) technology in the field of human-computer interaction is increasing. At present, the speech recognition technology mainly converts the speech signal to be recognized into text information through the signal processing module, the feature extraction module, the acoustic model, the language model (Language Model, LM), the dictionary and the decoder (Decoder), and completes the speech recognition.

In the process of speech recognition, the signal processing module and the feature extraction module first divide the speech signal to be recognized into a plurality of speech signal frames, and then enhance each speech signal frame by eliminating noise, channel distortion, etc., and then each speech. The signal frame is transformed from the time domain to the frequency domain and appropriate acoustic features are extracted from the converted speech signal frame. The acoustic model trained according to the characteristic parameters of the training speech library is input to the acoustic feature extracted by the feature extraction module, mapped to the pronunciation capable of describing the pronunciation feature of the speech signal frame, and the speech signal frame is mapped to each pronunciation. Probability, get the feature classification result.

The language model contains the associations between different words (such as words, words, phrases) and their probabilities (possibility), which are used to estimate the possibility of various textual information composed of different words. The decoder can establish an identification network based on the trained acoustic model, the language model and the dictionary, and identify each path in the network corresponding to various text information and the pronunciation of each text information, and then output the pronunciation for the acoustic model. Finding the best one path in the identification network, based on the path, the text information corresponding to the voice signal can be output with the maximum probability to complete the voice recognition.

However, the language model is generally based on a large number of corpus trained models, including the relationship and possibility between a large number of words. Therefore, the recognition network based on the speech model contains more nodes, and the number of branches per node is also very many. When the path search is performed in the identification network, the number of nodes involved in the pronunciation of each speech signal frame will increase exponentially, resulting in a large amount of path search, and the search process takes more time, which in turn reduces the speech recognition efficiency.

Summary of the invention

In view of this, the present invention provides a voice recognition method and apparatus, which solves the problem that the voice recognition process is time-consuming and effective. The problem of low rates.

According to a first aspect of the present invention, a speech recognition method is provided, comprising the steps of:

Acquiring a feature classification result of the speech signal to be identified; the feature classification result includes a pronunciation for describing a pronunciation feature of each speech signal frame and a probability that each speech signal frame is mapped to a corresponding pronunciation;

Filtering the pronunciations included in the feature classification result based on the probability contained in the feature classification result;

The speech signal is identified based on the filtered feature classification result.

In an embodiment, the filtering, according to a probability included in the feature classification result, filtering the pronunciation included in the feature classification result, including:

Determining whether the probability that any speech signal frame is mapped to the corresponding pronunciation satisfies a predetermined filtering rule;

If the corresponding pronunciation satisfies a predetermined filtering rule, the corresponding pronunciation is filtered out.

In one embodiment, if any of the speech signal frames are mapped to a probability of a corresponding pronunciation, and a probability difference between the maximum mapping probability of the speech signal frame, within a predetermined difference range, the corresponding pronunciation is determined. Meet predetermined filtering rules;

If the probability that any of the speech signal frames are mapped to the corresponding utterance is less than the probability that the speech signal frame is mapped to a predetermined number of utterances in the utterance, then the corresponding utterance is determined to satisfy the predetermined filtering rule.

In one embodiment, the predetermined number is any of the following:

The number of pronunciations retained in the feature classification result in the pronunciation corresponding to the frame speech signal frame;

The product of the predetermined ratio threshold and the total number of pronunciations corresponding to the frame of the speech signal frame.

In an embodiment, the filtering, according to a probability included in the feature classification result, filtering the pronunciation included in the feature classification result, including:

Obtaining a histogram distribution of the probability of mapping any speech signal frame to each pronunciation;

Obtaining a beam width corresponding to the histogram distribution;

Having a probability that the probability is distributed outside the beam width is determined to satisfy the pronunciation of the predetermined filtering rule;

The pronunciation that satisfies the predetermined filtering rule is deleted from the pronunciation included in the feature classification result.

In one embodiment, the removing the pronunciation that satisfies the predetermined filtering rule from the pronunciation included in the feature classification result includes:

If the probability that any of the speech signal frames are mapped to the corresponding pronunciation satisfies a predetermined filtering rule, the pronunciation is determined as a candidate pronunciation;

If any one of the adjacent speech signal frames of the predetermined number of frames of the speech signal frame, the probability of mapping to the candidate pronunciation satisfies a predetermined filtering rule, the candidate pronunciation is deleted from the pronunciation included in the feature classification result ;

If the adjacent speech signal frame of the predetermined number of frames of the speech signal frame does not satisfy the predetermined filtering rule, the candidate pronunciation is retained in the pronunciation included in the feature classification result.

According to a second aspect of the present invention, a speech recognition apparatus is provided, comprising:

a classification result obtaining module, configured to obtain a feature classification result of the speech signal to be identified; the feature classification result includes a pronunciation for describing a pronunciation feature of each speech signal frame and a probability that each speech signal frame is mapped to a corresponding pronunciation;

a pronunciation filtering module, configured to filter a pronunciation included in the feature classification result based on a probability included in the feature classification result;

a voice recognition module, configured to identify the voice signal based on the filtered feature classification result.

In an embodiment, the pronunciation filtering module further includes:

a first filtering module, configured to: when a probability difference between a probability of mapping any voice signal frame to a corresponding pronunciation and a maximum mapping probability of the voice signal frame, within a predetermined difference range, the corresponding pronunciation Filtering;

The second filtering module is configured to filter the corresponding pronunciation when the probability of mapping any of the speech signal frames to the corresponding pronunciation is less than the probability that the speech signal frame is mapped to a predetermined number of pronunciations.

In an embodiment, the pronunciation filtering module includes:

a probability distribution module, configured to obtain a histogram distribution of a probability that any speech signal frame is mapped to each pronunciation;

a beam width determining module, configured to acquire a beam width corresponding to the histogram distribution;

a pronunciation determining module, configured to determine a pronunciation whose probability is distributed outside the beam width, to determine a pronunciation that satisfies the predetermined filtering rule;

The pronunciation deletion module is configured to delete the pronunciation that satisfies the predetermined filtering rule from the pronunciation included in the feature classification result.

In an embodiment, the pronunciation filtering module includes:

a candidate pronunciation module, configured to determine the pronunciation as a candidate pronunciation when a probability that any of the speech signal frames are mapped to the corresponding pronunciation meets a predetermined filtering rule;

a candidate pronunciation deletion module, configured to map any one of adjacent speech signal frames of a predetermined number of frames of the speech signal frame When the probability of hitting the candidate pronunciation satisfies a predetermined filtering rule, the candidate pronunciation is deleted from the pronunciation included in the feature classification result;

a candidate pronunciation retaining module, configured to retain the candidate pronunciation in the feature classification result when the probability of mapping the candidate speech to the adjacent speech signal frame of the predetermined number of frames of the speech signal frame does not satisfy the predetermined filtering rule In the pronunciation.

Embodiments provided by the present invention, when identifying a voice signal, first acquiring a feature classification result of the voice signal, and then filtering the pronunciation included in the feature classification result based on a probability included in the feature classification result, then In the process of recognizing the speech signal, it is no longer necessary to perform an identification operation related to the filtered pronunciation, such as no need to search the identification network for the path related to the filtered pronunciation, thereby effectively reducing the time spent in the speech recognition process. In turn, the speech recognition efficiency can be improved.

DRAWINGS

1 is a flow chart showing a voice recognition method according to an exemplary embodiment of the present invention;

2 is a flow chart showing a voice recognition method according to another exemplary embodiment of the present invention;

FIG. 3 is a logic block diagram of a voice recognition apparatus according to an exemplary embodiment of the present invention; FIG.

FIG. 4 is a logic block diagram of a voice recognition apparatus according to another exemplary embodiment of the present invention; FIG.

FIG. 5 is a hardware configuration diagram of a voice recognition apparatus according to an exemplary embodiment of the present invention.

detailed description

Exemplary embodiments will be described in detail herein, examples of which are illustrated in the accompanying drawings. The following description refers to the same or similar elements in the different figures unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Instead, they are merely examples of devices and methods consistent with aspects of the invention as detailed in the appended claims.

The terminology used in the present invention is for the purpose of describing particular embodiments, and is not intended to limit the invention. The singular forms "a", "the" and "the" It should also be understood that the term "and/or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items.

It should be understood that although the terms first, second, third, etc. may be used to describe various information in the present invention, such information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, the first information may also be referred to as the second information without departing from the scope of the invention. Similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein may be interpreted as "when" or "when" or "in response to a determination."

In the speech recognition of the embodiment of the present invention, an acoustic model and a language model are involved in the recognition process, wherein the acoustic model is performed on differences in acoustics, phonetics, environmental variables, and gender, accent, and the like of the person who emits the voice. Knowledge representation, the LSTM (Long Short-Term Memory), CTC (Connectionist temporal classification) model, or Hidden Markov Model HMM can be used to train the speech contained in the training speech library to obtain the acoustics of the speech. The mapping of features to pronunciations constitutes an acoustic model that is related to the modeling unit. If the modeling unit is a syllable, the pronunciation is a syllable; if the modeling unit is a phoneme, the pronunciation is a phoneme; if the modeling unit is a state constituting a phoneme, the pronunciation is a state.

When training acoustic models, it is considered that the pronunciation will vary with words, speed, intonation, light accent, and dialect, etc. The training speech library needs to cover different words, speed, tone, and weight. A large number of voices, such as sounds and dialects, that influence the factors of pronunciation. In addition, considering the accuracy of speech recognition, a smaller pronunciation unit such as a syllable, a phoneme, a state, or the like can be selected as the modeling unit. Therefore, based on the large amount of speech contained in the training speech library and the predetermined modeling unit, the model training will construct a large number of acoustic models. In the process of speech recognition, a large number of acoustic models are used to classify the speech signals to be recognized, and the obtained feature classification results will contain a large number of pronunciations (categories), such as: 3,000 to 10,000 pronunciations.

In addition, the current speech recognition technology needs to recognize the text information corresponding to the speech signal, and it is necessary to search for all possible paths in the identification network for each pronunciation, and an index increment in the form of an index is generated in the search process. If you search for all possible paths involved in 3,000 to 10,000 pronunciations in the recognition network, the storage resources and calculations required for the search may exceed the limits of the speech recognition system. Therefore, the current speech recognition technology takes a lot of time. And resources, there is a problem of low efficiency of speech recognition, and the present invention proposes a solution for how to improve the efficiency of speech recognition.

In order to solve the problem of low speech recognition efficiency, the solution of the present invention improves the feature classification result obtained by the speech recognition process, and sets a filtering rule according to the device resource and the recognition efficiency requirement involved in the speech recognition, and then, when recognizing the speech signal, Firstly acquiring the feature classification result of the voice signal, and then filtering the pronunciation included in the feature classification result based on the probability included in the feature classification result, so that in the process of identifying the voice signal, there is no need to identify in the network Searching for the path associated with the filtered pronunciation, thus effectively reducing the time spent in the search process, thereby improving speech recognition efficiency. The speech recognition process of the present invention will be described in detail below with reference to the accompanying drawings.

Please refer to FIG. 1. FIG. 1 is a flowchart of a voice recognition method according to an exemplary embodiment of the present invention. The embodiment can be applied to various electronic devices having voice processing capabilities, and may include the following steps S101-S103:

Step S101: Acquire a feature classification result of the speech signal to be identified; the feature classification result includes a pronunciation for describing a pronunciation feature of each speech signal frame and a probability that each speech signal frame is mapped to a corresponding pronunciation.

Step S102: Filter the pronunciations included in the feature classification result based on the probability included in the feature classification result.

Step S103: Identify the voice signal based on the filtered feature classification result.

In the embodiment of the present invention, the voice signal may be a voice sent by a user collected by a local voice collection device in real time, or may be a voice transmitted remotely by the voice collection device. When acquiring the feature classification result of the speech signal, the speech signal may be pre-processed in real time through the speech preprocessing module in the field, and the pre-processed speech signal is extracted by the feature extraction module, and the extracted feature may include PLP ( Perceptual Linear Predictive, LPCC (Linear Predictive Cepstral Coefficient), FBANK (Mel-Scale Filter Bank), MFCC (Mel-Frequency Cepstral Coefficients, Meyer) The cepstrum coefficient is equal to, and then the extracted features are processed correspondingly by the acoustic model to obtain the feature classification result, and the probability contained in the feature classification result is used to indicate the possibility that the speech signal frame is mapped to the corresponding pronunciation. In other examples, it is also possible to directly receive the feature classification results transmitted by other terminal devices.

After obtaining the feature classification result, the solution of the present invention considers that part of the pronunciation included in the feature classification result has low correlation with the speech signal frame of the speech signal to be recognized, and has little influence on the speech recognition accuracy rate, and is reduced. When the influence of the large number of pronunciations included in the feature classification result on the speech recognition efficiency, the pronunciations having less influence on the speech recognition accuracy rate can be filtered out from the feature classification result before the speech recognition based on the feature classification result. The number of pronunciations included in the feature classification result is reduced, thereby improving the efficiency of speech recognition.

In general, the lower the correlation between the pronunciation and the speech signal frame to be recognized, the lower the probability that the speech signal frame is mapped to the speech when the acoustic characteristics of the speech signal are classified by the acoustic model. Therefore, the pronunciation included in the feature classification result may be filtered based on the probability that the speech signal frame is mapped to each pronunciation, and after filtering, the probability that any speech signal frame is mapped to the filtered pronunciation is smaller than the speech signal frame is mapped to other The probability of pronunciation.

In addition, when filtering the less relevant pronunciation, considering the need of the speech recognition accuracy rate of different application scenarios, it is also necessary to measure the influence of the filtered pronunciation on the speech recognition accuracy rate, and therefore, according to the speech recognition accuracy rate Requirements, various filtering rules that can limit the degree of influence of the filtered pronunciation on the recognition accuracy. For each predetermined filtering rule, when filtering the pronunciation included in the feature classification result, the probability that any speech signal frame is mapped to the corresponding pronunciation is Whether the predetermined filtering rule is satisfied, and if the corresponding pronunciation satisfies a predetermined filtering rule, the corresponding pronunciation is filtered out. Filtered pronunciation generally refers to the pronunciation that is removed from the feature classification results.

Here are a few ways to filter the pronunciations contained in the feature classification results:

Filtering method 1: filtering the low-probability pronunciation according to the predetermined number, the predetermined number may refer to the number of pronunciations in the pronunciation corresponding to the speech signal frame retained in the feature classification result; or may refer to the predetermined proportional threshold corresponding to the speech signal frame The product of the total number of pronunciations. At the time of filtering, if the probability that any of the speech signal frames are mapped to the corresponding utterance is less than the probability that the speech signal frame is mapped to a predetermined number of utterances in the utterance, it is determined that the corresponding utterance satisfies a predetermined filtering rule.

The predetermined proportional threshold may be set by the designer of the present invention according to the required speech recognition accuracy, for example, set to 1/4, which refers to the ratio of the retained pronunciation to the total number of pronunciations.

In an example, in actual filtering, the pronunciation may be deleted from the feature classification result in the order of probability from small to large, and when the number of reserved pronunciations is proportional to the number of original pronunciations, the predetermined proportional threshold is satisfied, and the feature is completed. Filtering of classification results.

In other examples, the predetermined ratio threshold may refer to the ratio of the number of unfiltered pronunciations to the filtered pronunciations. In the actual filtering, the pronunciation can be selected in the feature classification result according to the order of probability, and when the ratio of the number of selected pronunciations to the number of remaining pronunciations satisfies a predetermined proportional threshold, the feature classification result is completed. Filtering.

In practical applications, when the predetermined number refers to the number of pronunciations in the pronunciation corresponding to the frame speech signal frame that are retained in the feature classification result, the designer of the present invention can set the predetermined number according to the required speech recognition accuracy, for example, , set to any of the values from 2000 to 9000. When filtering, the pronunciations to which each speech signal frame is mapped may be arranged in descending order of probability, and then the pronunciations arranged in the predetermined number of bits are deleted from the feature classification result, and the result of the feature classification is completed. Filtering, the predetermined number of bits is equal to the predetermined number of values.

In other examples, the predetermined number may refer to the number of utterances that are not filtered out, for example, set to 1000. In actual filtering, the pronunciations to which each speech signal frame is mapped may be arranged in descending order of probability, and then the pronunciations arranged in the predetermined number of bits are retained in the feature classification result, and other pronunciations are characterized. The classification result is deleted, and the filtering of the feature classification result is completed, and the predetermined number of bits is equal to the value of the quantity threshold. In other embodiments, other technical means may also be used to perform filtering according to a pair of feature classification results in a filtering manner, which is not limited by the present invention.

Filtering method 2: filtering the low-probability pronunciation according to the predetermined difference threshold, which can be set by the designer of the present invention according to the required speech recognition accuracy, for example, set to -0.5, and the filtering is filtered. The probability difference between the probability of falling out and the pronunciation of the most probable probability that the same speech signal frame is mapped. When filtering, if the probability of any speech signal frame being mapped to the corresponding pronunciation, the probability difference between the maximum mapping probability of the speech signal frame and the predetermined difference range, Then determining that the corresponding pronunciation meets a predetermined filtering rule, and filtering the corresponding pronunciation.

In an example, in actual filtering, the pronunciations to which each speech signal frame is mapped may be arranged in descending order of probability, and the speech signal frame is mapped to the probability of the pronunciation arranged in the first position. Determined as the maximum probability, and then from the beginning of the pronunciation of the last bit, sequentially obtain the difference between the probability that the speech signal frame is mapped to each pronunciation and the maximum probability, and if the difference is less than -0.5, the pronunciation is classified from the feature. The result is removed. In other embodiments, other technical means may be used to filter the feature classification result according to the filtering method. The present invention does not limit this.

Filtering method 3: filtering the pronunciations distributed outside the beam width according to the histogram distribution of the probability. In actual filtering, the histogram distribution of the probability that any speech signal frame is mapped to each pronunciation may be acquired first; The map distribution corresponds to a beam width; then the pronunciation whose probability distribution is outside the beam width is determined to satisfy the pronunciation of the predetermined filtering rule; finally, the pronunciation of the predetermined filtering rule is satisfied, from the feature classification result The included pronunciation is removed. In practical applications, the beam width can be determined by the designer of the present invention according to the required speech recognition accuracy and the distribution of the histogram. For example, the preset needs to filter out 8000 low-probability pronunciations, which can be from the histogram. The low probability side starts to find 8000 pronunciations, and the 8000th pronunciation position is determined as the beam width boundary. In other embodiments, other technical means may be used to filter the feature classification result according to the filtering manner. The present invention does not limit this.

After filtering the pronunciations included in the feature classification result according to any of the above filtering methods, the predetermined identification network may be directly retrieved, and the path related to the pronunciation included in the filtered feature classification result is searched for the best. a path based on which the text information corresponding to the speech signal to be recognized is output with maximum probability to complete speech recognition. The identification network mentioned herein may refer to the speech signal to be recognized by the decoder according to the well-trained acoustics. Models, language models, and identification networks established by dictionaries.

When searching for the best path, the probability (acoustic score) contained in the feature classification result can be converted to be similar to the association probability (language score) between the words (such as words, words, phrases) contained in the phonetic model. The numerical space, and weighted addition, as a comprehensive score of the path search process, each speech signal frame is limited by a preset threshold value, and the difference from the optimal path is greater than the threshold value, then the The path is discarded, otherwise it is reserved. After each voice signal frame completes the search, all paths are sorted according to the preset maximum number of paths, and only the optimal path is retained until the last frame is completed. Path map.

In some examples, the modeling unit of the acoustic model that outputs the feature classification result is small. For example, the state is the modeling unit. Since a single phoneme can be composed of three to five states, the voice signal formed by the pronunciation of one phoneme can be Dividing into a plurality of speech signal frames, therefore, it is prone to a situation in which the acoustic characteristics of a plurality of consecutive speech signal frames are relatively similar, and the pronunciation of each frame in the continuous speech signal frames in the feature classification result is prone to a similar situation. . In response to this situation, if the present invention is based on the probability and predetermined filtering rules contained in the feature classification result, the pronunciations mapped to each frame of the speech signal frame are respectively Line filtering can easily filter out the pronunciation that has a great influence on the recognition accuracy. In order to avoid erroneous filtering of such pronunciations, the filtering status of continuous speech signal frames can be comprehensively considered when filtering the classification results. The specific implementation process can refer to the figure. The method shown in 2 includes the following steps S201-S205:

Step S201: Acquire a feature classification result of the speech signal to be identified; the feature classification result includes a pronunciation for describing a pronunciation feature of each speech signal frame and a probability that each speech signal frame is mapped to a corresponding pronunciation.

Step S202: If any of the speech signal frames are mapped to the probability of the corresponding pronunciation, the predetermined filtering rule is satisfied, and the pronunciation is determined as the candidate pronunciation.

Step S203: If any one of the adjacent speech signal frames of the predetermined number of frames of the speech signal frame, the probability of mapping to the candidate pronunciation satisfies a predetermined filtering rule, the candidate pronunciation is included in the feature classification result. Deleted in pronunciation.

Step S204: If the adjacent speech signal frame of the predetermined number of frames of the speech signal frame does not satisfy the predetermined filtering rule, the candidate pronunciation is retained in the pronunciation included in the feature classification result. .

Step S205: Identify the voice signal based on the filtered feature classification result.

In the embodiment of the present invention, the predetermined filtering rule may be any one of the filtering methods 1 to the filtering method 4 described above, and may also be other filtering rules that can limit the influence degree of the filtered pronunciation on the recognition accuracy.

The predetermined number of frames of consecutive speech signal frames can be set by the designer of the present invention according to the required speech recognition accuracy, for example, set to 6, adjacent first three frames and adjacent last three frames.

It can be seen from the above embodiment that the speech recognition method of the present invention first acquires the feature classification result of the speech signal, and then, based on the probability contained in the feature classification result, the pronunciation included in the feature classification result. By performing filtering, in the process of recognizing the voice signal, it is no longer necessary to perform the recognition operation related to the filtered pronunciation execution, such as no need to search the identification network for the path related to the filtered pronunciation, thereby effectively reducing the voice. Identifying the time spent in the process can improve speech recognition efficiency.

Furthermore, the voice recognition method in the embodiment of the present invention can be applied to human-computer interaction software of various electronic devices, for example, voice dialing, voice control, and voice search in a smart phone, and is applied to voice search in a smart phone if The user sends a voice in a predetermined range from the smartphone, and the voice recognition method applied to the voice search may obtain the feature classification result of the voice after receiving the user voice collected by the voice collection device, and then The probability included in the feature classification result, filtering the pronunciation included in the feature classification result, and then searching only the uncorrelated pronunciation-related path in the recognition network, and quickly identifying the text information corresponding to the user voice through the path search In turn, the voice assistant quickly responds to the user based on the recognition result.

Corresponding to the embodiment of the aforementioned method, the invention also provides an embodiment of the device.

Referring to FIG. 3, FIG. 3 is a logic block diagram of a voice recognition apparatus according to an exemplary embodiment of the present invention. The apparatus may include: a classification result obtaining module 310, a pronunciation filtering module 320, and a voice recognition module 330.

The classification result obtaining module 310 is configured to obtain a feature classification result of the speech signal to be identified; the feature classification result includes a pronunciation for describing a pronunciation feature of each speech signal frame, and each speech signal frame is mapped to a corresponding pronunciation. Probability.

The pronunciation filtering module 320 is configured to filter the pronunciations included in the feature classification result based on the probability included in the feature classification result.

The voice recognition module 330 is configured to identify the voice signal based on the filtered feature classification result.

In some examples, the pronunciation filtering module 320 can include:

a first filtering module, configured to: when a probability difference between a probability of mapping any voice signal frame to a corresponding pronunciation and a maximum mapping probability of the voice signal frame, within a predetermined difference range, the corresponding pronunciation Filter.

The second filtering module is configured to filter the corresponding pronunciation when the probability of mapping any of the speech signal frames to the corresponding pronunciation is less than the probability that the speech signal frame is mapped to a predetermined number of pronunciations.

In other examples, the pronunciation filtering module 320 may further include:

A probability distribution module is configured to obtain a histogram distribution of the probability that any speech signal frame is mapped to each pronunciation.

a beam width determining module for acquiring a beam width corresponding to the histogram distribution.

The pronunciation determining module is configured to determine a pronunciation whose probability is distributed outside the beam width as a pronunciation satisfying the predetermined filtering rule.

The pronunciation deletion module is configured to delete the pronunciation that satisfies the predetermined filtering rule from the pronunciation included in the feature classification result.

Referring to FIG. 4, FIG. 4 is a logic block diagram of a voice recognition apparatus according to another exemplary embodiment of the present invention. The apparatus may include a classification result acquisition module 410, a pronunciation filtering module 420, and a voice recognition module 430. The pronunciation filtering module 420 can include a candidate pronunciation determination module 421, a candidate pronunciation deletion module 422, and a candidate pronunciation retention module 423.

The classification result obtaining module 410 is configured to obtain a feature classification result of the speech signal to be identified; the feature classification result includes a pronunciation for describing a pronunciation feature of each speech signal frame, and each speech signal frame is mapped to a corresponding pronunciation. Probability.

The candidate pronunciation determining module 421 is configured to determine the pronunciation as a candidate pronunciation when the probability that any of the speech signal frames are mapped to the corresponding pronunciation meets a predetermined filtering rule.

The candidate pronunciation deleting module 422 is configured to: when any one of the adjacent speech signal frames of the predetermined number of frames of the speech signal frame is mapped to a predetermined filtering rule, the candidate pronunciation is from the feature. The pronunciation included in the classification result is deleted.

The candidate pronunciation retaining module 423 is configured to reserve the candidate pronunciation in the feature classification result when the probability of mapping the candidate speech to the adjacent speech signal frame of the predetermined number of frames of the speech signal frame does not satisfy the predetermined filtering rule. In the pronunciation contained.

The voice recognition module 430 is configured to identify the voice signal based on the filtered feature classification result.

The implementation process of the functions and functions of each unit (or module) in the above device is specifically described in the implementation process of the corresponding steps in the foregoing method, and details are not described herein again.

For the device embodiment, since it basically corresponds to the method embodiment, reference may be made to the partial description of the method embodiment. The device embodiments described above are merely illustrative, wherein the units or modules described as separate components may or may not be physically separate, and the components displayed as units or modules may or may not be physical units. Or modules, which can be located in one place, or distributed to multiple network units or modules. Some or all of the modules may be selected according to actual needs to achieve the objectives of the solution of the present invention. Those of ordinary skill in the art can understand and implement without any creative effort.

Embodiments of the speech recognition apparatus of the present invention can be applied to an electronic device. This can be implemented by a computer chip or an entity, or by a product having a certain function. In a typical implementation, the electronic device is a computer, and the specific form of the computer may be a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email transceiver device. , game consoles, tablets, wearables, Internet TVs, smart locomotives, driverless cars, smart refrigerators, other smart home devices, or a combination of any of these devices.

The device embodiment may be implemented by software, or may be implemented by hardware or a combination of hardware and software. Taking the software implementation as an example, as a logical device, a processor of the electronic device in which it is located reads a corresponding computer program instruction in a readable medium such as a non-volatile memory into a memory. From a hardware level, as shown in FIG. 5, is a hardware structure diagram of an electronic device in which the voice recognition device of the present invention is located, except for the processor, the memory, the network interface, and the nonvolatile memory shown in FIG. 5. The electronic device in which the device is located in the embodiment may also include other hardware according to the actual function of the electronic device, and details are not described herein again. The memory of the electronic device can be stored in a processor executable a program instruction; the processor may be coupled to the memory for reading the program instructions stored in the memory, and in response, performing the following operations: acquiring a feature classification result of the voice signal to be identified; the feature classification result is included for description The pronunciation of the pronunciation feature of each speech signal frame and the probability that each speech signal frame is mapped to the corresponding pronunciation; filtering the pronunciation contained in the feature classification result based on the probability contained in the feature classification result; based on the filtered The feature classification result identifies the speech signal.

In other embodiments, the operations performed by the processor may be referred to the related description in the foregoing method embodiments, and details are not described herein.

The above are only the preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalents, improvements, etc., which are made within the spirit and principles of the present invention, should be included in the present invention. Within the scope of protection.

Claims (10)

A speech recognition method, comprising the steps of: Acquiring a feature classification result of the speech signal to be identified; the feature classification result includes a pronunciation for describing a pronunciation feature of each speech signal frame and a probability that each speech signal frame is mapped to a corresponding pronunciation; Filtering the pronunciations included in the feature classification result based on the probability contained in the feature classification result; The speech signal is identified based on the filtered feature classification result. The method according to claim 1, wherein the filtering of the pronunciations included in the feature classification result based on the probability contained in the feature classification result comprises: Determining whether the probability that any speech signal frame is mapped to the corresponding pronunciation satisfies a predetermined filtering rule; If the corresponding pronunciation satisfies a predetermined filtering rule, the corresponding pronunciation is filtered out. The method of claim 2 wherein: If the probability difference between the probability of any speech signal frame being mapped to the corresponding pronunciation and the maximum mapping probability of the speech signal frame is within a predetermined difference range, determining that the corresponding pronunciation satisfies a predetermined filtering rule; If the probability that any of the speech signal frames are mapped to the corresponding utterance is less than the probability that the speech signal frame is mapped to a predetermined number of utterances in the utterance, then the corresponding utterance is determined to satisfy the predetermined filtering rule. The method of claim 3 wherein said predetermined number is any of the following: The number of pronunciations retained in the feature classification result in the pronunciation corresponding to the frame speech signal frame; The product of the predetermined ratio threshold and the total number of pronunciations corresponding to the frame of the speech signal frame. The method according to claim 1, wherein the filtering of the pronunciations included in the feature classification result based on the probability contained in the feature classification result comprises: Obtaining a histogram distribution of the probability of mapping any speech signal frame to each pronunciation; Obtaining a beam width corresponding to the histogram distribution; Acquiring a probability that is distributed outside the beam width is determined to satisfy a predetermined filtering rule; The pronunciation satisfying the predetermined filtering rule is deleted from the pronunciation included in the feature classification result. The method according to any one of claims 1 to 5, wherein the filtering of the pronunciation included in the feature classification result based on the probability contained in the feature classification result comprises: If the probability that any of the speech signal frames are mapped to the corresponding pronunciation satisfies a predetermined filtering rule, the pronunciation is determined as a candidate pronunciation; If any one of the adjacent speech signal frames of the predetermined number of frames of the speech signal frame, the probability of mapping to the candidate pronunciation satisfies a predetermined filtering rule, the candidate pronunciation is deleted from the pronunciation included in the feature classification result ; If the adjacent speech signal frame of the predetermined number of frames of the speech signal frame, the probability of mapping to the candidate pronunciation is not satisfied. The filter rule is retained, and the candidate pronunciation is retained in the pronunciation included in the feature classification result. A speech recognition device, comprising: a classification result obtaining module, configured to obtain a feature classification result of the speech signal to be identified; the feature classification result includes a pronunciation for describing a pronunciation feature of each speech signal frame and a probability that each speech signal frame is mapped to a corresponding pronunciation; a pronunciation filtering module, configured to filter a pronunciation included in the feature classification result based on a probability included in the feature classification result; a voice recognition module, configured to identify the voice signal based on the filtered feature classification result. The device according to claim 7, wherein the pronunciation filtering module further comprises: a first filtering module, configured to: when a probability difference between a probability of mapping any voice signal frame to a corresponding pronunciation and a maximum mapping probability of the voice signal frame, within a predetermined difference range, the corresponding pronunciation Filtering; The second filtering module is configured to filter the corresponding pronunciation when the probability of mapping any of the speech signal frames to the corresponding pronunciation is less than the probability that the speech signal frame is mapped to a predetermined number of pronunciations. The device according to claim 7, wherein the pronunciation filtering module comprises: a probability distribution module, configured to obtain a histogram distribution of a probability that any speech signal frame is mapped to each pronunciation; a beam width determining module, configured to acquire a beam width corresponding to the histogram distribution; a pronunciation determining module, configured to determine a pronunciation whose probability is distributed outside the beam width, to determine a pronunciation that satisfies the predetermined filtering rule; The pronunciation deletion module is configured to delete the pronunciation that satisfies the predetermined filtering rule from the pronunciation included in the feature classification result. The apparatus according to any one of claims 7 to 9, wherein the pronunciation filtering module comprises: a candidate pronunciation determining module, configured to determine the pronunciation as a candidate pronunciation when a probability that any of the speech signal frames are mapped to the corresponding pronunciation meets a predetermined filtering rule; a candidate pronunciation deleting module, configured to classify the candidate pronunciation from the feature when any one of adjacent frames of the predetermined number of frames of the voice signal frame is mapped to a predetermined filtering rule The result is deleted from the pronunciation; a candidate pronunciation retaining module, configured to retain the candidate pronunciation in the feature classification result when the probability of mapping the candidate speech to the adjacent speech signal frame of the predetermined number of frames of the speech signal frame does not satisfy the predetermined filtering rule In the pronunciation.

Images