JP3917880B2 - Speech recognition apparatus, speech recognition method, and speech recognition program - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a speech recognition apparatus for recognizing uttered speech.
[0002]
[Prior art]
Speech recognition technology has become widely used in recent years. With the development of computer technology in recent years, not only word speech recognition but also continuous speech recognition can be used with software on a PC.
[0003]
In speech recognition, a user may utter a word (unknown word) that is not stored in the recognition dictionary. As a method for dealing with unknown words in continuous speech recognition, a phonological typewriter is used as described in “Handling of Unknown Words in Continuous Speech Recognition” (Science Technical Report, SP91-96, Dec. 1991). There is something. The phonological typewriter is a model of voice characteristics using subwords so that all possible utterances in Japanese can be recognized, and phonemes are often used for the subwords.
[0004]
Hereinafter, the conventional unknown word processing using the phoneme typewriter will be described. FIG. 3 is a block diagram showing an example of conventional unknown word processing using a phoneme typewriter.
[0005]
The voice of the speaker input to the speech recognition device 300 is input to the input unit 301 and converted into a digital signal. The voice waveform converted into a digital signal is input to the acoustic analysis unit 302 and analyzed. As an analysis method, a short-time spectrum analysis method is often used in which a relatively short time window is taken every 20 to 40 msec and the window is shifted every 8 to 16 msec. The speech waveform cut out by the time window is converted into a time series of feature vectors for each unit called a frame having the cut out time length. The feature vector is obtained by extracting the feature amount of the voice spectrum at that time, and is usually 10 to 50 dimensions, and a mel frequency cepstrum coefficient or the like is widely used. The converted feature vector is output to the recognition unit 303.
[0006]
For the acoustic model 307, HMM (Hidden Markov Model) prepared for each recognition unit is widely used, and a phoneme piece is often used as the recognition unit. The HMM is a nondeterministic stochastic finite automaton having a plurality of states, and is used as a statistical signal source model in which an unsteady signal source is represented by a connection of stationary signal sources. The HMM is expressed as a collection of states that transition, and includes a transition probability that represents the probability of state transition and an output probability that outputs the probability of the observation vector when the state transitions. An acoustic model expressed in HMM used for speech recognition models the characteristics of phonemes. In this HMM, the state roughly corresponds to a phonological event (stable interval such as closing, bursting, friction, steady vowel). The output probability is the probability of fluctuation of the signal output with the transition. For each word included in the vocabulary of the recognition dictionary, a series (hereinafter referred to as an HMM output series) constituted by outputs in a state corresponding to each recognition unit (phoneme etc.) is an input signal series (each feature vector). Is obtained from the output probability and the transition probability, and the word having the maximum value is used as the recognition result, thereby realizing speech recognition. In this case, the sequence of the input signal is obtained as observation data of an already occurring event, and it is determined how much the output sequence of the HMM is plausible compared with that data. In order to determine the correctness of a hypothesis (HMM output sequence) that explains an already occurring event (input signal sequence), an analysis is performed using the concept of likelihood instead of probability. When calculating the likelihood of an HMM state sequence, the sum of log likelihoods is often obtained instead of the product of likelihoods. The Viterbi algorithm is widely used as an algorithm for obtaining such a state sequence. In the case of continuous speech recognition, there are grammatical restrictions on each word in the sentence according to the language model.
[0007]
Parameters such as output probability and transition probability are learned in advance by an algorithm called a Baum-Welch algorithm by giving corresponding learning speech. In the following, it is assumed that an HMM whose recognition unit is a phoneme is stored in the acoustic model 307.
[0008]
The recognition dictionary 304 stores information on recognizable vocabulary. Word notation and phoneme symbol string are stored.
[0009]
As the language model 305, an n-gram model based on the vocabulary included in the recognition dictionary 304 is widely used. The n-gram model is a kind of statistical language model that estimates probability from sample data by a statistical method. Regarding the implementation of the language model using the n-gram model, “speech recognition system” (Ohm Publishing Co., Ltd.) Bureau).
[0010]
The phonological chain model 306 stores phonological chain rules that may be in Japanese. In unknown word processing using a phoneme typewriter, phoneme recognition for unknown word detection is performed in parallel with normal continuous speech recognition. The phoneme chain model 306 is often stored in the same data structure as the language model 305 in order to make the recognition unit 303 common.
[0011]
The recognition unit 303 performs speech recognition processing using the information of the output of the acoustic analysis unit 302, the acoustic model 307, the language model 305, and the phonological chain model 306. As for normal continuous recognition processing, a method for realizing it is described in the above "voice recognition system" (Ohm Publishing Co.). However, here, in the recognition process, it is assumed that the search is performed by a one-pass frame synchronization beam search. At this time, phoneme recognition using a phoneme typewriter in each frame is also performed in parallel. FIG. 4 is a flowchart showing processing in each frame in the recognition unit 303. Hereinafter, the recognition process in the i-th frame will be described with reference to FIG.
[0012]
First, in step 401, normal recognition processing is performed on the corresponding frame. Specifically, the likelihood of each hypothesis is calculated from the feature vector output from the acoustic analysis unit 302 using the acoustic model 307 and the language model 305, and the hypothesis having a low likelihood is removed from the evaluation target (pruning). Process. The likelihood calculation method and pruning are also described in the above "speech recognition system" (Ohm Publishing House). Furthermore, the cumulative likelihood of each hypothesis is stored. For example, the cumulative likelihood corresponding to the frame i is set as an array P1 [i], and P1 is stored separately for each hypothesis.
[0013]
Next, in step 402, the same processing as in step 401 is performed using the phoneme chain model 306 instead of the language model 305. However, the maximum likelihood in each hypothesis in step 401 is set as a threshold value for the pruning score. Because there is a linguistic restriction in the case of recognition using the language model 305, the maximum value is higher in the likelihood in recognition using the phonological chain model 306, which may be a phonological chain that can be in Japanese. Furthermore, for unknown word detection, it is only necessary to have one recognition hypothesis using the phonological chain model. At this time, the maximum cumulative likelihood when the phoneme chain model in the corresponding frame is used is stored in the array P2 [i].
[0014]
When the above processing is completed for all the frames, for each word of the sentence corresponding to the hypothesis having the highest score, an analysis using the phonological chain model of the last frame in each word (here, j-th) is performed. A difference between the value stored in the phonological chain model weight storage unit and the cumulative likelihood in the analysis using the language model 305 is obtained as S, and stored as S in each word. That is, if the value stored in the phonological chain model weight storage unit is α, α × P2 [j] −P1 [j] = S. This value is obtained for each word. Finally, the notation of each word of the sentence corresponding to the hypothesis with the highest score and the value of S are output to the unknown word section detection unit 309.
[0015]
The unknown word section detection unit 309 receives the output of the recognition unit 303 as an input. A word string having S greater than 0 in each input word is regarded as an unknown word, and a character string obtained by converting the notation of the word to “unknown word” is output to the output unit 310. Here, the phoneme chain model weight storage unit 308 stores phoneme chain model weights appropriately determined in advance in consideration of the speech environment such as noise.
[0016]
The output unit 310 is a device that can output a character string, such as a display, and outputs the character string output from the unknown word section detection unit.
[0017]
Since the above phoneme chain model weight greatly varies depending on the speaker and the utterance environment, in order to detect unknown words with high accuracy, it is necessary to change the phoneme chain model weight depending on the environment. To detect unknown words that do not depend on the difference in environment, using the method found in Japanese Patent Laid-Open No. 4-255900, an appropriate phoneme chain can be determined from the likelihood of the analysis by the phonetic typewriter and the analysis with language restrictions. There is a method for obtaining model weights.
[0018]
[Problems to be solved by the invention]
However, the method described in the above prior art has a problem that the appropriate phoneme chain model weight varies depending on the environment. In the technique found in Japanese Patent Laid-Open No. 4-255900 for solving this problem, there is a disadvantage that an appropriate phonological chain model weight is not known unless it is analyzed to the end of the utterance. I do not know at all whether or not. In speech recognition using a statistical language model such as n-gram, the recognition result after an unknown word is uttered is not very reliable, and even if only a portion having a large difference from a phonological typewriter is recognized as an unknown word. In many cases, the correct word is not selected as a result in the subsequent part, so there is a problem in recognizing a part other than the unknown word.
[0019]
SUMMARY OF THE INVENTION An object of the present invention is to provide a speech recognition apparatus capable of detecting unknown words with high accuracy and realizing unknown word processing that is robust to recognition of parts other than unknown words.
[0020]
[Means for Solving the Problems]
The present invention relates to a speech recognition apparatus having a recognition dictionary storing words, a language model in which the words stored in the recognition dictionary have been learned in advance, and a phoneme chain model storing phoneme chain rules that can be in the recognition target language. A recognition unit that performs an analysis using the language model and an analysis using the phonological chain model, a phonological chain model weight storage unit that stores a value for weighting the analysis result using the phonological chain model, A phoneme chain model weight changing unit that changes a value stored in the phoneme chain model weight storage unit using a result of an analysis using a language model and an analysis using the phoneme chain model. And
[0021]
The language model weight storage unit stores the weight of the language model, and the language model weight storage unit stores the result of the analysis using the language model and the analysis using the phonological chain model. A language model weight changing unit for changing the weights.
[0022]
In the present invention, the language model weight changing unit may calculate a difference between likelihoods obtained by analyzing the weight of the language model using the language model in the recognition unit and the analysis using the phonological chain model, or The value is obtained by a function of values stored in the language model weight storage unit.
[0023]
Further, according to the present invention, the phoneme chain model weight storage unit is a value of a difference in likelihood obtained by analyzing the phoneme chain model weight using the language model in the recognition unit and the analysis using the phoneme chain model. Or it changes to the value obtained by the function of the value memorize | stored in the said phoneme chain model weight memory | storage part.
[0024]
Further, according to the present invention, the recognition unit searches for a plurality of paths in the search at the time of likelihood calculation, and the language model weight change unit is stored in the weight storage unit of the language model in the second and subsequent paths. It is characterized by maintaining the value.
[0025]
Further, according to the present invention, the recognition unit searches for a plurality of paths in the search at the time of likelihood calculation, and the phoneme chain model weight changing unit stores the phoneme chain model weight storage unit in the second and subsequent paths. It is characterized in that the value being maintained is maintained.
[0026]
Further, according to the present invention, the language model weight storage unit separately stores a value before utterance, and the language model weight change unit stores the value in the language model weight storage unit when it is determined that the utterance is finished. The recorded value is changed to a value before utterance.
[0027]
Further, in the present invention, the phonological chain model weight storage unit separately stores a value before utterance, and the phonological chain model weight change unit stores the phonological chain model weight storage when it is determined that the utterance is finished. The value stored in the section is changed to a value before utterance.
[0028]
Further, the present invention is characterized in that the language model weight changing unit maintains a value stored in the language model weight storage unit in a processing unit determined to be a silent part.
[0029]
Further, the present invention is characterized in that the phonological chain model weight changing unit maintains a value stored in the phonological chain model weight storage unit in a processing unit determined to be a silent part.
[0030]
In addition, the present invention uses a recognition dictionary that stores words, a language model that learns each word stored in the recognition dictionary in advance, and a phoneme chain model that stores rules of phoneme chains that may be in the recognition target language. In the speech recognition method, for each processing unit divided by time, a recognition unit that performs analysis using the language model and analysis using the phonological chain model, and outputs a maximum likelihood value in each analysis; A phonological chain model weight storage means for storing a value for weighting the analysis using the phonological chain model, and a maximum of each of the analysis using the language model and the analysis using the phonological chain model for each processing unit. Phonological chain model weight changing means for changing the value stored in the phonological chain model weight storage means by using the likelihood value.
[0031]
Further, the present invention provides language model weight storage means for storing the weight of the language model, and the maximum likelihood of each of the analysis using the language model and the analysis using the phonological chain model for each processing unit. Language model weight changing means for changing the weight stored in the language model weight storage means by using a value.
[0032]
The present invention also includes a recognition dictionary that stores words, a language model that learns each word stored in the recognition dictionary in advance, and a phonological chain model that stores rules of phonological chains that may be in the recognition target language. In the speech recognition method, a recognition means for performing an analysis using the language model and an analysis using the phonological chain model, and a phonological chain model weight storage means for storing a value for weighting the analysis result using the phonological chain model And phonological chain model weight changing means for changing the value stored by the phonological chain model weight storage means using the results of the analysis using the language model and the analysis using the phonological chain model. It is characterized by that.
[0033]
Further, the present invention provides a language model weight storage means for storing the weight of the language model, and is stored by the language model weight storage means from the results of the analysis using the language model and the analysis using the phonological chain model. Language model weight changing means for changing the weight.
[0034]
The present invention also provides a recognition dictionary that stores word information, a language model that learns each word stored in the recognition dictionary in advance, and a phonological chain model that stores rules of phonological chains that can exist in the recognition target language. A speech recognition program that uses a computer to recognize a means for performing an analysis using the language model and an analysis using the phonological chain model, and to store a value that weights the analysis result using the phonological chain model Phoneme chain model weight storage means for performing, and using the results of the analysis using the language model and the analysis using the phoneme chain model, the phoneme chain model weight for changing the value stored by the phoneme chain model weight storage means A speech recognition program for functioning as a changing means is provided.
[0035]
According to the present invention, in the above speech recognition program, the computer is obtained from the results of the analysis using the language model weight storage means for storing the weight of the language model, the analysis using the language model, and the analysis using the phonological chain model. Provided is a speech recognition program for functioning as a language model weight changing means for changing a weight stored by a language model weight storage means.
[0036]
DETAILED DESCRIPTION OF THE INVENTION
Hereinafter, the present invention will be described in detail based on examples.
[0037]
FIG. 1 is a block diagram showing an example of a speech recognition apparatus according to the present invention. The speech recognition apparatus 100 includes an input unit 101, an acoustic analysis unit 102, a recognition unit 103, an acoustic model 107, a recognition dictionary 104, a language model 105, a phonological chain model 106, a phonological chain model weight changing unit 110, and a phonological chain model weight storage. Unit 111, language model weight change unit 108, language model weight storage unit 109, unknown word section detection unit 112, and output unit 113.
[0038]
The input unit 101 converts the input voice into a digital signal.
[0039]
The acoustic analysis unit 102 converts the input digital signal into a feature vector time series.
[0040]
The acoustic model 107 stores phonological features of utterances using an HMM prepared for each phoneme unit. In order to support the two-pass search in the recognition unit 103, two models are stored: a model with low accuracy but capable of performing faster recognition, and a model with higher accuracy but recognized at low speed. Yes.
[0041]
The recognition dictionary 104 stores a character string representing a word notation and a character string representing a phoneme symbol string as recognizable vocabulary information.
[0042]
The language model 105 stores statistical language information based on the n-gram model.
[0043]
The phoneme chain model 106 stores phoneme chain rules that may be in Japanese with the same data structure as the language model.
[0044]
The phoneme chain model weight storage unit 111 stores phoneme chain model weights used for recognition in the recognition unit, and 1 is stored as an initial value. How to use the phoneme chain model weight will be described later.
[0045]
The language model weight storage unit 109 stores the weight of the language model used in the recognition process in the recognition unit, and 1 is stored as an initial value.
[0046]
The recognition unit 103 performs speech recognition processing using the information of the output of the acoustic analysis unit 102, the acoustic model 107, the language model 105, and the phonological chain model 106. Here, it is assumed that the recognition unit 103 performs recognition processing in two passes and performs frame-synchronized beam search in both passes. However, in the first pass, candidates are narrowed down while speeding up with a simple acoustic model, and two passes. High-accuracy recognition is performed using high-accuracy acoustic models with the eyes. FIG. 2 is a flowchart showing the processing of the i-th frame in the first pass in the recognition unit. Hereinafter, the recognition process in the i-th frame will be described with reference to FIG.
[0047]
In step 201, normal recognition processing is performed in the corresponding frame using the acoustic model 107 and the language model 105 in the same manner as the method described in the related art. Here, the cumulative likelihood corresponding to the frame i is set as the array P1 [i], and the likelihood P1 is stored for each hypothesis.
[0048]
Next, in step 202, the same processing as in step 201 is performed using the phonological chain model 106 instead of the language model 105. At this time, the maximum likelihood of each hypothesis in step 201 is set as a threshold value of the score for pruning. Similarly, the maximum cumulative likelihood when the phoneme chain model in the corresponding frame is used is stored as an array P2 [i].
[0049]
In step 203, it is determined whether it is a silent part. In the acoustic model 107, it is determined that there is silence when the HMM state corresponding to silence has the highest likelihood. If there is no sound, the process proceeds to the next step. If not, the process proceeds to step 204. This is because, in the silent part, there is no difference in likelihood in recognition using a language model, and therefore the expected effect cannot be obtained even if the value of the language model weight or the phoneme chain model weight is changed. The change notification, the value of the array P2 [i], and the value of the array P1 [i] are output to the phoneme chain model weight change unit.
[0050]
In step 205, it is determined whether or not a notification for changing the value is sent to the language model weight changing unit. Here, assuming that the value stored in the phoneme chain model weight storage unit is α, when α × P2 [i] −P1 [i] is a positive value, in step 206, the array P2 [i] The value and the value of the array P1 [i] are output to the language model weight change unit and the phonological chain model weight storage unit together with the change notification. Otherwise, go to the next frame.
[0051]
In the second pass, almost the same processing is performed except that a high-accuracy acoustic model is used. However, in the second pass, the language model weight is not notified to the language model weight change unit and the phonological chain model weight change unit. The changing unit and the phonological chain model weight changing unit maintain the previous values. This is because the analysis using the phonological chain model in all input frames is completed in the first pass, so that all information of the utterance can be used.
[0052]
Further, the language model weight value and the phonological chain model weight changed by the current utterance do not necessarily have to be returned to the values before the utterance. When applying to devices such as mobile phones where the environment such as noise during utterance is not specified, this value was memorized even if the environment changed significantly by the next occurrence by returning to the value before utterance Can prevent adverse effects. On the other hand, when applying to a device such as a home computer where the environment at the time of utterance is almost constant, the value changed by the current utterance is stored in the next utterance without returning to the state before utterance. Can be recognized with an appropriate value.
[0053]
When the above processing is completed for all frames and the second pass processing is also completed, the unknown word section detection unit 112 outputs a hypothesis having the maximum likelihood and a difference S between the cumulative likelihoods of each word of the hypothesis.
[0054]
The phoneme chain model weight changing unit 110 receives the notification from the recognition unit 103 and changes the value stored in the phoneme chain model weight storage unit. Here, in the analysis using the phoneme chain model, there is no linguistic restriction by the recognition dictionary or the language model, so that the likelihood obtained is increased. When comparing the likelihood of the analysis using the phonological chain model and the analysis using the language model, the likelihood obtained by the analysis using the phonological chain model becomes large even if the utterance is not an unknown word. Therefore, the coefficient indicating the weight (penalty) is always set to a value smaller than 1.
[0055]
The change is made α = (P2 [i] −P1 [i]) / i, where α is a value stored in the phoneme chain model weight storage unit. This is because the difference in likelihood between the array P2 [i] and the array P1 [i] due to the difference in environment often increases due to the difference in the characteristics of the acoustic model 107 and the spoken voice.
[0056]
The language model weight change unit 108 changes the value stored in the language model weight storage unit in accordance with the notification from the recognition unit 103. The change is β = β × ((P2 [i] −P1 [i]) / P2 [i]) where β is a value stored in the language model weight storage unit. This is because the larger the difference between the values of the array P2 [i] and the array P1 [i], the less reliable the numerical value of the language model after that.
[0057]
The unknown word section detection unit 112 receives the output of the recognition unit 103 as an input, converts a word having a cumulative likelihood difference S greater than 0 in each input word as an unknown word, and converts the word notation to “unknown word” The character string is output to the output unit 113.
[0058]
The output unit 113 uses a device that can output a character string, such as a display, and outputs the character string output from the recognition unit 103.
Example 1
Hereinafter, a specific processing operation when the sentence “go to Jiyugaoka” is uttered will be shown. At this time, the word “Jiyugaoka” is not included in the recognition dictionary 104.
[0059]
First, the input voice is converted into a digital signal by the input unit 101. The voice converted into a digital signal is divided into frames by a short-time spectrum analysis method, and the voice in each frame is converted into a time series of feature vectors by the acoustic analysis unit 102 and output to the recognition unit 103.
[0060]
In the recognition unit 103, processing is performed in each frame according to the flowchart of FIG.
[0061]
First, in step 201, the likelihood P1 is calculated using the language model 105 and the acoustic model 107.
[0062]
In step 202, the likelihood P2 is calculated using the phoneme chain model 106 and the acoustic model 107.
Here, since the word “Jiyugaoka” is not included in the recognition dictionary 104, the maximum likelihood P1 is another word included in the recognition dictionary 104, such as “self-employed”. Since “Jiyugaoka” and “self-employed” have different phoneme arrangements, the likelihood P1 value of “self-employed” is high among the words included in the recognition dictionary 104, but the restriction of phoneme arrangement by the recognition vocabulary The likelihood P2 calculated using the unacceptable phoneme chain model 106 is considered to be lower.
[0063]
Next, in step 203, it is determined whether or not there is a silent part. If there is no sound, the process proceeds to the next frame as it is. If not, the process proceeds to step 204.
[0064]
In step 204, a weight change notification and the values of likelihood P1 and likelihood P2 are output to the phoneme chain model weight change unit. This is because the appropriate phonological chain model weight varies greatly depending on the surrounding environment such as noise and reverberation, and therefore the appropriate phonological chain model weight is set based on the values obtained in the previous analysis.
[0065]
Next, in step 205, it is determined whether to change the weight of the language model. If α × P2 [i] −P1 [i] is a positive value, the weight of the language model is changed in step 205. Otherwise, the process proceeds to the next frame without doing anything. In the second pass, almost the same processing is performed except that a high-accuracy acoustic model is used. In the second pass, the language model weight is not notified to the language model weight change unit and the phonological chain model weight change unit. The changing unit and the phonological chain model weight changing unit maintain the previous values. When the processing is completed for all the frames, the language model weight value and the phonological chain model weight are returned to the values before utterance.
[0066]
Since the language model 105 stores statistical linguistic information based on the n-gram model, words that cannot be used in Japanese as “going to self-employment” are highly likely to be obtained using the acoustic model 107. However, the likelihood obtained using the language model 105 is low, and the likelihood of the word sequence such as “self-employed”, “operating self-employed” is higher, and “go to” in the utterance ", A recognition error occurs. However, by changing the weight of the language model in step 205, the linguistic restrictions are relaxed, and the word sequence such as “go to” with high acoustic likelihood can be continued after “self-employed”. can do. When the above processing is completed for all frames and the second pass processing is also completed, the unknown word section detection unit 112 is notified of the hypothesis with the maximum likelihood (in this example, “go to self-employment”) and each word of the hypothesis. The cumulative likelihood difference S is output. When the weight of the language model is not changed, the hypothesis with the maximum likelihood is often a correct word string in Japanese, such as “self-employed”.
[0067]
In the unknown word section detection unit 112, the output of the recognition unit 103 is the hypothesis “go to self-employment” and the likelihood difference S for each of the words “self-employment”, “ni”, and “go”. As an input. Here, a word whose cumulative likelihood difference S in each word is larger than 0 is determined as an unknown word. From the difference in phoneme arrangement between “Jiyugaoka” and “Self-employed”, the difference between the likelihood P1 and the likelihood P2 is large in “Self-employed”, so that “Self-employed” can be determined as an unknown word. If the phonological chain model weight is not changed, it is difficult to set an appropriate phonological chain model weight depending on the utterance environment, so “self-employed” cannot be determined as an unknown word, or “ni” or “go” is unknown It may be judged as a word. The unknown word section detection unit 112 converts the “self-employed” part to “unknown word” and outputs it to the output unit 113, and the output unit 113 outputs the result to a display device such as a display.
(Example 2)
Hereinafter, as an example of the case where the speech recognition apparatus using the speech recognition method of the present invention is used as a word speech recognition apparatus, an operation when the word “Jiyugaoka” is uttered is shown. FIG. 5 shows an example of a speech recognition apparatus that performs word recognition using the speech recognition method of the present invention. In the case of recognition processing for only words, statistical language information by n-gram is not used. Also, in word recognition, it is often the case that a plurality of paths often used in the previous sentence recognition are not recognized. Here, it is assumed that processing is performed in one pass. Furthermore, the word “Jiyugaoka” is not included in the recognition dictionary 504.
[0068]
First, the input voice is converted into a digital signal by the input unit 501. The voice converted into a digital signal is divided into frames by a short-time spectrum analysis method, and the voice in each frame is converted into a time series of feature vectors by the acoustic analysis unit 502 and output to the recognition unit 503.
[0069]
In the recognition unit 503, processing is performed in each frame according to the flowchart of FIG.
[0070]
First, in step 601, the likelihood P1 is calculated using the acoustic model 507.
Next, in step 602, likelihood P2 is calculated using the phoneme chain model 506 and the acoustic model 507.
Here, since the word “Jiyugaoka” is not included in the recognition dictionary 504, the maximum likelihood P1 is another word included in the recognition dictionary 504, such as “self-employed”. Since “Jiyugaoka” and “self-employed” have different phoneme arrangements, the likelihood P1 value of “self-employed” is high among the words included in the recognition dictionary 504, but the restriction of phoneme arrangement by the recognition vocabulary is limited. It is often lower than the likelihood P2 calculated using the unacceptable phoneme chain model 506.
[0071]
Next, in step 603, it is determined whether or not there is a silent part. If there is no sound, the process proceeds to the next frame as it is. If not, the process proceeds to step 604.
[0072]
Next, in step 604, a weight change notification and likelihood P1 and likelihood P2 values are output to the phonological chain model weight changing unit. This is because the appropriate phonological chain model weight varies greatly depending on the surrounding environment such as noise and reverberation, and therefore the appropriate phonological chain model weight is set based on the values obtained in the previous analysis. When the processing is completed for all frames, the phonological chain model weight is returned to the value before utterance.
[0073]
In the unknown word section detection unit 512, each of the candidate “self-employed” output from the recognition unit 503 and the cumulative likelihood difference S of “self-employed” are input. Here, a word with S greater than 0 is determined as an unknown word. From the difference in phoneme arrangement between “Jiyugaoka” and “Self-employed”, the difference between the likelihood P1 and the likelihood P2 is large in “Self-employed”, so that “Self-employed” can be determined as an unknown word. If the phoneme chain model weight is not changed, it may be difficult to determine “self-employed” as an unknown word because it is difficult to set an appropriate phoneme chain model weight depending on the utterance environment. In the unknown word section detection unit 512, the “self-employed” part is converted to “unknown word” and output to the output unit 513, and the output unit 513 outputs the result to a display device such as a display.
[0074]
【The invention's effect】
As described above, according to the present invention, unknown words can be detected with high accuracy when performing speech recognition. In addition, it is possible to realize unknown word processing that is robust even in recognition of parts other than unknown words.
[Brief description of the drawings]
FIG. 1 is a block diagram showing an example of a speech recognition apparatus according to the present invention.
FIG. 2 is a flowchart illustrating an operation of a recognition unit according to the present invention.
FIG. 3 is a block diagram showing an example of a conventional speech recognition apparatus.
FIG. 4 is a flowchart showing the operation of a recognition unit in a conventional speech recognition apparatus.
FIG. 5 is a block diagram showing an example of a speech recognition apparatus that performs word recognition using the speech recognition method according to the present invention.
FIG. 6 is a flowchart showing an operation of a recognition unit according to the present invention for word recognition.
[Explanation of symbols]
100, 300, 500 Speech recognition device
101, 301, 501 input section
102, 302, 502 Acoustic analysis unit
103, 303, 503 recognition unit
104, 304, 504 recognition dictionary
105,305 Language model
106, 306, 506 Phonological chain model
107, 307, 507 Acoustic model
108 Language model weight change section
109 Language model weight storage
110, 510 Phonological chain model weight change unit
111, 308, 511 Phoneme chain model weight storage unit
112, 309, 512 Unknown word section detector
113, 310, 513 output section

Claims (16)

In a speech recognition apparatus having a recognition dictionary that stores words, a language model that has learned words stored in the recognition dictionary in advance, and a phonological chain model that stores phonological chain rules that may be in the recognition target language,
A recognition unit that performs analysis using the language model and analysis using the phonological chain model;
A phoneme chain model weight storage unit for storing a value for weighting the analysis result using the phoneme chain model;
A phonological chain model weight changing unit that changes a value stored in the phonological chain model weight storage unit using a result of the analysis using the language model and an analysis using the phonological chain model. A featured voice recognition device. A language model weight storage unit for storing the weight of the language model, and a language for changing the weight stored in the language model weight storage unit from the results of the analysis using the language model and the analysis using the phonological chain model The speech recognition apparatus according to claim 1, further comprising a model weight changing unit. The language model weight change unit includes a likelihood difference value obtained by analyzing the language model weight in the recognition unit using the language model and the analysis using the phonological chain model, or the language model weight storage unit. The speech recognition apparatus according to claim 1, wherein the value is obtained by a function of a value stored in The phonological chain model weight storage unit is a likelihood difference value obtained by analyzing the phonological chain model weight in the recognition unit using the language model and an analysis using the phonological chain model, or the phonological chain model weight. The voice recognition device according to claim 1 or 2, wherein the voice recognition device is changed to a value obtained by a function of a value stored in the storage unit. The recognition unit searches for a plurality of paths in the search at the time of likelihood calculation, and the language model weight change unit maintains the value stored in the weight storage unit of the language model in the second and subsequent paths. The speech recognition apparatus according to claim 1 or 2, wherein The recognizing unit searches a plurality of paths in the search at the time of likelihood calculation, and the phoneme chain model weight changing unit maintains a value stored in the phoneme chain model weight storage unit in the second and subsequent paths. The speech recognition apparatus according to claim 1 or 2, wherein The language model weight storage unit stores a value before utterance separately, and the language model weight change unit determines the value stored in the language model weight storage unit before utterance when it is determined that the utterance has ended. The voice recognition device according to claim 1 or 2, wherein the value is changed to a value of. The phonological chain model weight storage unit separately stores a value before utterance, and the phonological chain model weight change unit stores the value stored in the phonological chain model weight storage unit when it is determined that the utterance has ended. The speech recognition apparatus according to claim 1, wherein the value is changed to a value before utterance. The speech recognition according to claim 1 or 2, wherein the language model weight changing unit maintains a value stored in the language model weight storage unit in a processing unit determined to be a silent part. apparatus. 3. The phonological chain model weight changing unit maintains a value stored in the phonological chain model weight storage unit in a processing unit determined as a silent part. Voice recognition device. In a speech recognition method using a recognition dictionary that stores words, a language model in which each word stored in the recognition dictionary is learned in advance, and a phonological chain model that stores phonological chain rules that can be recognized in the recognition target language,
Recognizing means for performing analysis using the language model and analysis using the phonological chain model for each processing unit divided by time, and outputting a maximum likelihood value in each analysis,
Phoneme chain model weight storage means for storing a value for weighting the analysis using the phoneme chain model;
A phonological chain model that changes the value stored in the phonological chain model weight storage means by using the maximum likelihood values of the analysis using the language model and the analysis using the phonological chain model for each processing unit. And a weight changing means. Language model weight storage means for storing the weight of the language model;
Language model weight change for changing the weight stored in the language model weight storage means using the maximum likelihood values of the analysis using the language model and the analysis using the phonological chain model for each processing unit The speech recognition method according to claim 11, further comprising: means. In a speech recognition method comprising: a recognition dictionary storing words; a language model in which each word stored in the recognition dictionary has been learned in advance; and a phonological chain model storing rules of phonological chains that may exist in a recognition target language;
Recognition means for performing analysis using the language model and analysis using the phonological chain model;
Phoneme chain model weight storage means for storing a value for weighting the analysis result using the phoneme chain model;
Phonological chain model weight changing means for changing the value stored by the phonological chain model weight storage means using the results of the analysis using the language model and the analysis using the phonological chain model. A feature of speech recognition. Language model weight storage means for storing the weight of the language model;
The language model weight changing means for changing the weight stored by the language model weight storage means from the result of the analysis using the language model and the analysis using the phonological chain model, 14. The speech recognition method according to 13. A speech recognition program that uses a recognition dictionary that stores word information, a language model that learns each word stored in the recognition dictionary in advance, and a phonological chain model that stores phonological chain rules that may exist in the recognition target language. There,
Computer
Recognition means for performing analysis using the language model and analysis using the phonological chain model;
Phoneme chain model weight storage means for storing a value for weighting the analysis result using the phoneme chain model;
Speech recognition for functioning as a phonological chain model weight changing means for changing the value stored by the phonological chain model weight storage means using the analysis using the language model and the analysis using the phonological chain model program. Computer
Language model weight storage means for storing the weight of the language model;
16. The function according to claim 15, wherein the function is used as a language model weight changing means for changing the weight stored by the language model weight storage means from the result of the analysis using the language model and the analysis using the phonological chain model. Speech recognition program.

Images