JP4037709B2 - Speech recognition method and speech recognition system - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a speech recognition method and a speech recognition system , and more particularly to a speech recognition method and speech recognition system using speaker adaptation technology.
[0002]
[Prior art]
In order to improve speech recognition performance for a specific speaker, the speech recognition apparatus uses a speaker adaptation technique for learning the speech of the specific speaker. Such speaker adaptation techniques can be broadly classified into two types: supervised and unsupervised. The teacher here refers to a phoneme notation string representing the utterance content of the input utterance. The supervised adaptation is an adaptation method in the case where the phoneme notation sequence for the input utterance is known, and it is necessary to indicate the utterance vocabulary to the unknown speaker in advance at the time of adaptation.
[0003]
On the other hand, unsupervised adaptation is an adaptation method when the phoneme notation sequence for the input utterance is unknown, and does not limit the utterance content of the input utterance to the unknown speaker, that is, the utterance content for the unknown speaker. There is no need to give an instruction, and the adaptation can be performed without making the unknown speaker aware of the input speech that is actually being used for speech recognition.
In general, unsupervised adaptation has a lower recognition performance after adaptation than supervised adaptation, and thus supervised adaptation is often used.
[0004]
Hereinafter, a speech recognition apparatus using a conventional supervised adaptation technique will be described with reference to FIG.
The speaker's utterance S input to the speech recognition device 1 is input to the input pattern creation unit 2 and undergoes a process such as AD conversion and speech analysis, and a feature vector for each unit called a frame having a certain length of time. Converted to series. This time series of feature vectors is referred to herein as an input pattern. The feature vector is an extracted feature of the speech spectrum at that time, and is usually 10 to 100 dimensions.
[0005]
On the other hand, the standard pattern storage unit 3 stores a hidden Markov model (HMM). The HMM is one of voice information source models, and its parameters can be learned using a speaker's voice. The HMM is usually prepared for each predetermined recognition unit. Here, a phoneme is taken as an example of the recognition unit. Therefore, the phoneme HMM is stored in the standard pattern storage unit 3 as a standard pattern as shown in FIG. As the phoneme HMM, for example, an unspecified speaker HMM learned in advance using the utterances of many speakers is used.
Assume that 1000 words are to be recognized, that is, a correct answer for one word is obtained from 1000 word recognition candidates. At the time of word recognition, the vocabulary pattern creation unit 5 creates an HMM (word MFF) of recognition candidate words by connecting the HMMs of the phonemes constituting the word. In the case of 1000 word recognition, a word HMM for 1000 words is created. That is, as shown in FIG. 4, the vocabulary pattern creation unit 5 includes a word storage unit 5a that stores 1000 words and a word HMM creation unit 5b that creates a word HMM for each word by connecting phoneme HMMs. Yes.
[0006]
The recognition unit 4 recognizes an input pattern using the 1000-word word HMM created by the vocabulary pattern creation unit 5. The phoneme HMM is a model of a voice information source, and a statistical idea is introduced in the description of the standard pattern in order to deal with various fluctuations of the voice pattern. A phoneme HMM is usually composed of 1 to 10 states and state transitions therebetween. Normally, a start state and an end state are defined, and a symbol is output from each state and a state transition is performed every unit time. The phoneme speech pattern is represented as a time series of symbols output during the state transition from the start state to the end state. Symbol appearance probability is defined for each state, and transition probability is defined for each transition between states. The transition probability parameter is a parameter for expressing temporal fluctuation of the voice pattern. The appearance probability parameter represents the voice color fluctuation of the voice pattern. It is possible to calculate the occurrence probability when the input speech is assumed to be generated from a certain HMM by setting the probability of the start state to a certain value and multiplying the input pattern by the appearance probability and the transition probability for each state transition. it can.
[0007]
In speech recognition by HMM, word HMMs are prepared for 1000 recognition candidate words, and when speech is input, the occurrence probability is obtained in the word HMM of each recognition candidate word, and the maximum word HMM is generated. The recognition candidate word corresponding to the word HMM is determined as a recognition result. The recognition result word is sent to the recognition result output unit 6. The recognition result output unit 6 performs processing such as outputting the recognition result on the screen or sending a control command corresponding to the recognition result to another apparatus. In the above, the HMM is stored in the standard pattern storage unit 3 according to the phoneme, but the word HMM can also be stored according to the word as shown in FIG. In such a case, the vocabulary pattern creation unit 5 is unnecessary.
[0008]
Next, a supervised speaker adaptation technique for the speech recognition apparatus 1 will be described. In supervised speaker adaptation, a word to be uttered is instructed to a user in advance, and phoneme HMM parameters are updated using word notation and input speech. This is called supervised adaptation in the sense that the correct word for the utterance is known in advance.
[0009]
First, as in the case of recognition, the input pattern creation unit 7 creates an input pattern from the input voice. In the case of supervised adaptation, since the correct word is known in advance, the adaptation dictionary creation unit 8 creates an adaptation dictionary from the inputted correct word notation (input speech notation). Next, the vocabulary pattern creation unit 9a of the supervised adaptation unit 9 inputs the phoneme sequence of the adaptation dictionary and the adaptation initial HMM for each phoneme stored in advance in the adaptation initial standard pattern storage unit 9b. A word HMM corresponding to the pattern is created. Then, the adaptation unit 9c performs likelihood calculation between the input pattern and the adaptation word HMM, performs an adaptation process on one or a plurality of input patterns, and then calculates an average vector after adaptation. An HMM after adaptation is obtained, and the adapted HMM is input to the standard pattern storage unit 3 and stored instead of the standard HMM used so far.
The supervised adaptation technique described above is well known and is described in detail, for example, in Japanese Patent Laid-Open No. 7-230295.
[0010]
[Problems to be solved by the invention]
However, the supervised adaptation method has a drawback in that a word instructed from the apparatus must be uttered as training separately from the generation of speech in the original speech recognition, and the burden is large. In other words, it is necessary to work for a purpose (improving the recognition performance of speech recognition) different from the original purpose of installing the speech recognition device to improve the human interface (to make it easier to operate the device), which is complicated. In addition, there is a problem that imposes a burden on the user.
Accordingly, an object of the present invention is to provide a speech recognition method and a speech recognition system that can automatically have the same performance as a supervised adaptation method without making the user aware of it.
Another object of the present invention is to provide a speech recognition method and a speech recognition system capable of providing a performance equivalent to that of a supervised adaptation method with a simple configuration.
[0011]
[Means for Solving the Problems]
A first aspect of the present invention is a speech recognition method that compares registered speech patterns of a plurality of words with an input speech pattern and recognizes a speech input word based on the most similar registered speech pattern. A speech recognition engine for recognizing the input speech using the standard pattern ( speech pattern ) registered in the speech recognition device, and the speech recognition result by the speech recognition engine and the input speech pattern are externally registered speech A step of sending to a pattern improvement device; a step of recognizing speech based on an input speech pattern sent from the speech recognition device in the registered speech pattern improvement device; A step of considering as a teacher signal in learning, the teacher signal and the input speech pattern sent from the speech recognition device; And performing supervised speaker adaptation processing, comparing the teacher signal with the speech recognition result sent from the speech recognition apparatus to obtain different phonemes, and obtaining the improved speech data of the registered speech pattern of the different phonemes. A step of generating based on the result of the user adaptation processing, a step of transmitting the improved data of the registered voice pattern to the voice recognition device, and the voice recognition device being registered in the voice recognition engine by the sent improvement data. Updating the registered voice pattern .
[0012]
The second of the present invention is a speech recognition system, speech recognition word that is the voice input based on the registered voice patterns by comparing the plurality of words registered voice pattern and the input speech pattern, most similar a recognition device, provided with a registered voice pattern improved apparatus for improving the registered voice pattern in the voice recognition apparatus, the speech recognition apparatus, speech of the input speech using the reference pattern (voice patterns) registered in correspondence with the phoneme A speech recognition engine for recognizing the speech recognition result by the speech recognition engine and an input speech pattern are sent to an external registered speech pattern improvement device, and the registered speech pattern improvement device is sent from the speech recognition device Speech recognition is performed based on the input speech pattern, and the input speech notation that is the speech recognition result is used as a teacher signal in supervised speaker adaptive learning. None, a supervised speaker adaptation process is performed using the teacher signal and the input speech pattern sent from the speech recognition device, and the teacher signal and the speech recognition result sent from the speech recognition device are Comparing different phonemes, generating improved data of the registered speech pattern of the different phonemes based on the speaker adaptation processing result, transmitting the improved data of the registered speech pattern to the speech recognition device, and The recognizing device updates the registered speech pattern registered in the speech recognition engine with the received improved data .
According to the present invention, it is possible to perform supervised speaker adaptive learning of the speech recognition device in the underground, so that the user does not need to supply a teacher signal to the speech recognition device, and the learning itself is supervised. Since it is learning, there is no risk of learning going in the wrong direction, and highly accurate learning is possible and the recognition rate can be improved.
[0013]
DETAILED DESCRIPTION OF THE INVENTION
(A) Configuration of Speech Recognition System FIG. 1 is a configuration diagram of a speech recognition system including a speech recognition device according to the present invention, and the speech recognition device 11 attached to a device controlled by speech input improves recognition performance. As described above, the improvement center 12 for updating the parameters specifying the standard pattern stored in the standard pattern storage unit of the speech recognition apparatus 11 is provided. The voice recognition device 11 and the improvement center 12 can freely communicate with each other by any communication method, for example, wireless communication using a LAN or a mobile phone.
[0014]
In the voice recognition device 11, the voice input microphone 21 detects and outputs the voice input by the speaker, the AD converter 22 converts the voice signal into digital, and the voice recognition engine 23 uses the conventional example of FIG. The input speech is recognized with the same configuration as the recognition device 1). The voice data file / recognition result storage unit 24 stores the recognition result of the input voice and stores the input voice pattern data in the WAV format. The data transmission unit 25 transmits one or more recognition results stored in the storage unit 24 and the input voice pattern data to the improvement center as a pair. The data receiving unit 26 receives the improved data (parameters for specifying the standard pattern) from the improved center 12, and the installer 27 updates the standard pattern data of the speech recognition engine 23 with the received improved data.
[0015]
In the improvement center 12, the data receiving unit 31 receives a plurality of sets of recognition result / input pattern data sent from the speech recognition device 11, and the teacher signal generation unit 32 is received from the speech recognition area engine 23 of the speech recognition device 11. It has a much higher performance speech recognition engine, performs speech recognition based on each set of input pattern data sent, and outputs the recognition result as input speech notation (teacher signal) of the input pattern. The improvement unit 33 performs an adaptation process similar to that of the supervised adaptation unit 9 in FIG. 2 using the input speech notation) and the input pattern data sent from the speech recognition device 11, and the phoneme HMM that has been adapted Ask for. Thereafter, the recognition result sent from the speech recognition apparatus is compared with the teacher signal (input speech notation) to obtain different phonemes, and an HMM adapted to the different phonemes is output. The improved data storage unit 34 stores the improved data (HMM adapted to each phoneme) output from the improving unit 33, and the data transmitting unit 35 transmits the improved data to the speech recognition apparatus.
[0016]
(B) Overall operation of the speech recognition system (1) When the user uses the speech recognition device 11, the speech content (speech pattern) at that time and the recognition result recognized by the speech recognition engine 23 are each a speech data file and recognition. The result is recorded in the result storage unit 24. The utterance content is held in, for example, a WAV file format, and an ID code is given. The recognition result is recorded in, for example, a text format corresponding to the ID code.
(2) When a certain period of time (which may be set by the user or set by the manufacturer) elapses, the data transmission unit 25 determines whether the audio data file recorded in the audio data file / recognition result storage unit 24 The recognition results are paired and transmitted to the improvement center 12.
[0017]
(3) In the improvement center 12, the data receiving unit 31 stores the information received from the voice recognition device 11 in the built-in memory and sequentially sends the utterance contents and the recognition result to the teacher signal generation unit 32 and the improvement unit 33. input. The teacher signal generation unit 32 performs voice recognition based on the input voice pattern data, outputs the recognition result as an input voice notation of the input voice pattern, and the improvement unit 33 outputs the input voice notation and the input voice pattern data. A phoneme HMM is obtained by performing an adaptation process using the same. Thereafter, the improvement unit 33 compares the recognition result sent from the speech recognition apparatus 11 with the teacher signal (input speech notation) to obtain different phonemes, and outputs an adapted HMM of the different phonemes. The improved data storage unit 34 stores the improved data output from the improving unit 33 (HMM adapted to each phoneme), and the data transmitting unit 35 transmits the improved data to the speech recognition apparatus 11.
[0018]
(4) Upon receiving the improved data, the installer 27 installs the improved data (the HMM adapted for each phoneme) in the standard pattern storage unit in the speech recognition engine 23, and stores it instead of the standard HMM used so far. .
In the above, the standard HMM is a phoneme HMM, but it may be a word HMM.
[0019]
【The invention's effect】
As described above, according to the present invention, it is possible to perform supervised speaker adaptive learning of the speech recognition apparatus in the underground, so that it is not necessary for the user (speaker) to supply a teacher signal to the speech recognition apparatus.
Further, according to the present invention, since the learning itself is supervised learning, there is no risk of learning going in the wrong direction, and high-accuracy learning is possible, so that each user uses a speech recognition device. The recognition performance that can be enjoyed by the user can be improved.
Further, according to the present invention, the user's voice is sent to the center side equipped with a high-performance voice recognition engine to generate improved data, and the improved data is installed in the voice recognition engine of the user's voice recognition device. It is possible to provide a speech recognition apparatus with an improved recognition rate for unspecified speakers. In addition, the user's voice recognition engine does not require a high-performance engine, and can be inexpensive and have a small device configuration.
[Brief description of the drawings]
FIG. 1 is a configuration diagram of a voice recognition system including a voice recognition device of the present invention.
FIG. 2 is a speech recognition apparatus using a conventional supervised adaptation technique.
FIG. 3 is an example of storing a phoneme HMM as a standard pattern.
FIG. 4 is a configuration diagram of a vocabulary pattern creation unit.
FIG. 5 is an example of storing a word HMM as a standard pattern.
[Explanation of symbols]
DESCRIPTION OF SYMBOLS 11 Voice recognition apparatus 12 Improvement center 21 Microphone 22 for voice input AD converter 23 Voice recognition engine 24 Voice data file and recognition result storage part 27 Installer 31 Data reception part 32 Teacher signal generation part 33 Improvement part 34 Improvement data storage part

Claims (2)

In a speech recognition method for comparing a registered speech pattern of a plurality of words and an input speech pattern and recognizing a speech input word based on the most similar registered speech pattern,
It provided a voice recognition engine to the speech recognition of the input speech using the reference pattern (voice patterns) registered in correspondence with the phoneme to the speech recognition device, an external speech recognition result by the voice recognition engine and the input speech pattern To the registered voice pattern improvement device
In the registered voice pattern improving apparatus,
Voice recognition based on the input voice pattern sent from the voice recognition device,
The input speech notation that is the speech recognition result is regarded as a teacher signal in supervised speaker adaptive learning,
Supervised speaker adaptation processing is performed using the teacher signal and the input speech pattern sent from the speech recognition device, and the teacher signal is compared with the speech recognition result sent from the speech recognition device. Obtaining different phonemes, and generating improved data of registered speech patterns of the different phonemes based on the speaker adaptation processing results,
Transmitting the improved data of the registered voice pattern to the voice recognition device;
In the voice recognition device, and it updates the registered voice pattern by improving data has the sent and registered in the voice recognition engine,
A speech recognition method characterized by the above. A speech recognition device that compares a registered speech pattern of a plurality of words and an input speech pattern, recognizes a speech input word based on the most similar registered speech pattern, and improves the registered speech pattern in the speech recognition device In a speech recognition system provided with a registered speech pattern improvement device,
The speech recognition device includes a speech recognition engine that recognizes speech of an input speech using a standard pattern ( speech pattern ) registered in correspondence with phonemes, and obtains a speech recognition result and an input speech pattern by the speech recognition engine. Sent to an external registered voice pattern improvement device,
The registered speech pattern improvement device recognizes speech based on the input speech pattern sent from the speech recognition device, regards the input speech notation that is the speech recognition result as a teacher signal in supervised speaker adaptive learning, Supervised speaker adaptation processing is performed using the teacher signal and the input speech pattern sent from the speech recognition device, and the teacher signal is compared with the speech recognition result sent from the speech recognition device. Obtaining different phonemes, generating improved data of the registered speech pattern of the different phonemes based on the speaker adaptation processing result, and transmitting the improved data of the registered speech pattern to the speech recognition device,
The voice recognition device updates a registered voice pattern registered in the voice recognition engine with the improved data sent.
A speech recognition system characterized by that .

Images