JP3513284B2 - Voice recognition method and apparatus - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition method and its apparatus.

[0002]

2. Description of the Related Art In recent years, practical use of voice recognition systems has become popular. In practical use of voice recognition, a technique for specifying what was input is also very important, but a technique for determining that the input cannot be accepted by the voice recognition system is also very important.
In the speech recognition system using the hidden Markov model (hereinafter, HMM), in order to solve this problem, (1) a phoneme chain-driven phoneme HMM chain likelihood-based normalization method for the likelihood of the input speech and (2) language A method using a Garbage Model HMM for search processing is known as a general and effective method.

In the method (1), the output probability of the phoneme HMM used for speech recognition and its transition probability are calculated according to "description of a phoneme chain (for example, consonant + vowel or vowel + vowel chain is repeated)". On the other hand, in the method (2), an HMM in which a set of inputs desired to be rejected is learned in advance is created, and the HMM is calculated according to each state of the HMM.

[0004]

However, in the above-mentioned conventional method (1), a language search process and a search space (trellis or Viterbi Search and trel) for obtaining a speech recognition candidate are used.
Apart from the lis space), search processing and search space for phoneme chain drive (eg trellis, Viterbi search and trellis
Space).

Further, in the conventional method (2), the output probability calculation of the Garbage Model is further required in addition to the recognition unit HMM, and the Garbage Mod added to the language search processing is required.
A language search space for el and its processing are required. In either method, the search process for rejects (trellis or
Viterbi search) and its search space (trellis space)
The need for a special output probability calculation such as Garbage Model is a problem from the viewpoint of calculation amount and memory capacity.

The present invention eliminates the above-mentioned conventional drawbacks, eliminates the search processing for rejects (trellis or Viterbi search), and further reduces the search space for rejects (trellis space) to achieve compactness, high speed and high performance. Provided are a voice recognition method and a device for realizing voice recognition.

[0007]

[0008]

[Means for Solving the Problem ] To solve this problem
In the speech recognition method of the present invention, in the speech recognition method for recognizing speech using a hidden Markov model, the input speech is acoustically analyzed to obtain speech parameters, and the output probability for each state of each hidden Markov model is calculated. And the process of
A language search is performed using linguistic information such as a dictionary or grammar, and output probabilities of hidden Markov models and their states, and recognition candidates of the input speech and its likelihood are obtained, and output probabilities for each state of each hidden Markov model From the calculated results, the maximum state output probability and its state are obtained for the analyzed voice parameter, and the reject is determined by the state transition method.
Determining a likelihood for rejecting a voice input based on the state transition probability for the voice input, and determining whether to output or reject the recognition result of the voice input based on a threshold based on the ratio of the two likelihoods. It is characterized by having.

Here, in the determination of the state transition probability determined by the way of transition of the state showing the maximum state output probability,
The transition probability is given by an empirical method. Further, in the determination of the state transition probability determined by the way of transition of the state showing the maximum state output probability, all states of the total hidden Markov model are decomposed, and all states are freely changed while fixing the inside that determines the output probability. The transition probability is learned from the voice data from which the hidden Markov model is learned by giving an initial value as much as possible.

The speech recognition apparatus of the present invention is a speech recognition apparatus for recognizing speech using a hidden Markov model, and means for acoustically analyzing input speech to obtain speech parameters, and analyzed speech parameters. A means for calculating the output probability for each state of each hidden Markov model of, and a language search using the language information such as a dictionary or grammar and the output probability of the hidden Markov model and that state, and the recognition candidates of the input speech and its From the result of calculating the likelihood and the output probability for each state of each hidden Markov model, the means for obtaining the maximum state output probability and its state for the analyzed speech parameter, and the transition of the obtained state transition means for determining a likelihood for rejection of the speech entered by the state transition probability for rejection determined by how, on the basis of the ratio of the two likelihood, the threshold And having a means for determining whether to reject outputs the recognition result of the speech input Ri.

Here, the means for obtaining the likelihood for rejecting gives a transition probability by an empirical method, and determines the state transition probability by the way of transition of the state showing the maximum state output probability. Further, the means for obtaining the likelihood for rejecting is to decompose all the states of the total hidden Markov model, give an initial value so that all the states can freely transition while fixing the inside that determines the output probability, and then make a hidden Markov model again. The transition probability is learned from the voice data that has learned the model, and the state transition probability is determined by the way of transition of the state showing the maximum state output probability.

[0012]

DETAILED DESCRIPTION OF THE INVENTION

<Concept of speech recognition method of this embodiment> In the speech recognition method of this embodiment, each state of the phoneme HMM used for speech recognition is disassembled into pieces, and the state showing the maximum output probability is changed with respect to the input frame. Then, the transition probability can be determined independently of the transition probability of the phoneme recognition HMM. However, if the likelihood is simply calculated so that the maximum value transits, the reject likelihood r is always larger than the likelihood R of the recognition result, so a threshold th is introduced, and when the recognition target is input, R
/ R> th, and R / r ≦ when input is outside the recognition target
A threshold th that is th is introduced.

By such a method, when the speech to be recognized is input, the likelihood R of the sequence of recognition HMMs constituting the recognition target for the input shows a large value, while for the reject Likelihood r
Is also a large value, but R / r> th for the above threshold
And the recognition is done correctly. On the other hand, if a voice that is not the target of recognition is input,
Recognition HM that constitutes all recognition targets for that input
Likelihood R of the M sequence shows a small value, while likelihood r for rejection has a relatively large value for all speeches, so that R / r ≦ th for the above threshold. hand,
Rejected.

<Example of Configuration of Speech Recognition Device According to this Embodiment>
FIG. 1 is a diagram showing a configuration example of a voice recognition device according to the present embodiment. In the figure, 101 is a microphone or A / D for inputting voice.
Includes a voice input unit, 102 an acoustic analysis unit for obtaining a voice parameter, 103 an output probability calculation unit for calculating an output probability, 104 an HMM for calculating an output probability, 105
Is a language search unit for performing language processing, 106 is a grammar / dictionary used for language processing, 107 is a maximum output probability calculation unit for reject, 108 is a state transition probability calculation unit for reject, 10
Reference numeral 9 is a reject determination unit, and 110 is a display unit for outputting the result.

The speech recognition apparatus may be realized by operating a general-purpose computer with a specific program, or may be configured to use a special computer such as a neuro computer. In the case of a general-purpose computer, a program or special data (table, etc.) for realizing the present speech recognition method may be loaded from an external storage medium or by communication into the main memory and executed.

<Processing Flow of Speech Recognition Apparatus According to this Embodiment> A speech recognition apparatus configured by the above elements is shown in FIG.
It operates according to the process flow shown in. Voice input unit 201
The sound extracted by (corresponding to 101) is the acoustic analysis unit 2
In 02 (corresponding to 102), the voice parameter is analyzed for each frame, and the output probability calculating unit 203 (corresponding to 103) calculates the output probability using the HMM 204 (corresponding to 104). Language search unit 205 (corresponding to 105)
Is the output probability from the output probability calculation unit 203 and the HMM 20.
4 and the linguistic information of the grammar / dictionary 206 (corresponding to 106), a language search is performed, and the recognition candidate and its likelihood R (2
05-1) is output.

On the other hand, the maximum output probability calculation unit for reject 2
In 07 (corresponding to 107), the maximum output probability and its state are obtained for each frame of the input speech, and the state transition probability calculation unit for reject 208 (corresponding to 108) outputs the maximum probability, the transition probability between states. The likelihood r (208-1) for reject is calculated based on Finally, the likelihood R of the recognition candidate is compared with the likelihood r for rejection, and the rejection determination unit 209 (corresponding to 109) determines whether to output or reject the recognition result for the input voice (R / R>
th or ≤th), each result (recognition result 210-1,
The reject 210-2) is output.

FIG. 3 is an image diagram of the process for obtaining the rejection likelihood r. As shown in FIG. 3, in the speech recognition method according to the present embodiment, each state of the phoneme HMM used for speech recognition is disassembled into pieces, and the state showing the maximum output probability is transitioned to the input frame. The transition probability is determined independently of the transition probability of the phoneme recognition HMM.

Generally, the HMM likelihood calculation is performed by taking the logarithm and calculating the sum of log probability. Therefore, the following two can be considered regarding the state transition probability.

(1) In the empirical method, when the state does not change (transition to the same state), log (0.95)
When changing to a different state, a method of adding log (0.05) can be considered. In this case, the state transition probability table (more accurately, not the probability) is as follows.

[0021]

[Table 1] (2) Moreover, all states of all HMMs are decomposed, initial values are given so that transitions between all states can be freely performed while fixing the internals that determine output probabilities, and the state transition probabilities are again obtained by speech data that has learned HMMs. In the method of learning, the following table (example) of state transition probabilities is created as a result of learning, and a method of using these logarithmic values can be considered. For learning, an EM algorithm, which is a normal learning method for HMM, can be used. At this time, the parameter related to the output probability is fixed.

[0022]

[Table 2] <Processing example by the voice recognition device of the present embodiment> Hereinafter, the result of the voice recognition according to the present embodiment and the result of the voice recognition according to the conventional method will be compared based on a specific example.

FIG. 4 is a diagram showing a simulation of the reject process according to the present embodiment. Here, it is assumed that the words that can be input are “Osaka” and “Tokyo”, the inputs are “Osaka” and “Hokkaido”, and the reject threshold th is 0.90. When the input is “Osaka”, the respective likelihoods are obtained in the search sections of the recognition targets of “Tokyo” and “Osaka”, and the larger likelihood is the likelihood R of the recognition result. In this example, 950.0 of "Osaka" is the likelihood R, while the reject likelihood r is 1000.0. Therefore,
Since R / r = 0.95 and th = 0.90, R / r> th, and the input is not rejected, and the recognition result “Osaka” is output.

On the other hand, when the input is “Hokkaido”, the respective likelihoods are obtained in the search sections of “Tokyo” and “Osaka” which are recognition symmetry, and the larger likelihood is the likelihood of the recognition result. It becomes R. In this example, 800.0 of "Tokyo" is the likelihood R, while the reject likelihood r is 1000.0. Therefore, R / r = 0.85
Since and th = 0.90, R / r <th, and the input is rejected and the recognition result is not output.

FIG. 5 is a diagram showing a simulation of the reject process according to the conventional method. Here, the left side is the case where the phoneme chain is used, and the right side is the case where the Garbage Model (GB model) is used. Phoneme chain of the conventional method and Garbage Model
Then, we need the memory of the search space for the total number of states of the model used there, and the Viterbi or trellis calculation of the interval.

From the comparison between the present embodiment and the conventional example,
In this embodiment, only one state of memory is required, and Vite
It can be seen that the likelihood for rejection is obtained by a simple calculation that does not require rbi calculation or trellis calculation, but in addition to selection of the maximum output probability state and a simple relation between the maximum output probability and the state transition probability. It should be noted that although the HMM in which the phoneme is used as a unit is used in the present embodiment, there is no problem in any unit, and the present invention includes these. Further, the present invention may be applied to a system including a plurality of devices or an apparatus including one device. Further, it goes without saying that the present invention can be applied to the case where it is achieved by supplying a program to a system or an apparatus. In this case, the storage medium storing the program according to the present invention constitutes the present invention. Then, by reading the program from the storage medium to the system or device, the system or device
It works in a predetermined way.

[0027]

EFFECTS OF THE INVENTION By replacing the output calculation for reject with a simple maximum value calculation, each state of the phoneme HMM used for speech recognition is disassembled into pieces, and the state showing the maximum output probability for the input frame is calculated. Since the transition probability for the reject can be determined independently of the transition probability of the phoneme recognition HMM, a search process for the reject (trellis or Viterbi se) which is necessary in the conventional method is performed.
arch) and search space (trelli
The s space), reduced to the memory for holding the the <br/> accumulated likelihood (one value) for reject obtaining in Qian <br/> transfer probability of the output probabilities and their use reject the maximum value, compact possible to provide a method and apparatus for realizing high-speed high-performance speech recognition.

[Brief description of drawings]

FIG. 1 is a configuration diagram of a voice recognition device of the present embodiment.

FIG. 2 is a diagram showing a flow of processing in the voice recognition device in the present embodiment.

FIG. 3 is a diagram showing an image of a process for obtaining a rejection likelihood r according to the present embodiment.

FIG. 4 is a diagram showing a simulation of reject processing according to the present embodiment.

FIG. 5 is a diagram showing a simulation of a conventional reject process.

─────────────────────────────────────────────────── ─── Continuation of the front page (56) References JP-A-8-6588 (JP, A) Patent 3315565 (JP, B2) Watanabe, Tsukada, Rejection of unknown utterance by likelihood correction using voice recognition, electronic IEICE Transactions DII, Japan, December 1992, Vol. J75-D-II, No. 12, p. 2002-2009 Ozeki, The mutual information as a focusing function for speech recognition, Technical Report of IEICE [Voice], Japan, December 15, 1995, SP 95-101, p. 53-60 Noda, Sagayama, HMM-LR speech recognition by A * beam search using forward likelihood, IEICE technical report [Speech], Japan, June 17, 1994, SP94-23, p. 1-7 (58) Fields investigated (Int.Cl. ⁷ , DB name) G10L 15/00-15/28 JISST file (JOIS)

Claims (6)

(57) [Claims] 1. A speech recognition method for recognizing speech using a hidden Markov model, the method comprising: acoustically analyzing input speech to obtain speech parameters; calculating output probabilities for each state of each hidden Markov model; Language search using linguistic information such as grammar and grammar, output probability of hidden Markov model and its state, and recognition candidates of input speech and its likelihood are calculated, and output probability for each state of each hidden Markov model is calculated. Based on the result, the maximum state output probability and its state are obtained for the analyzed voice parameter, and the reject is determined by the state transition method.
Determining a likelihood for rejecting a voice input based on a state transition probability for a voice, and determining whether to output or reject a recognition result of a voice input based on a threshold value based on a ratio of the two likelihoods. A method for recognizing speech, comprising: 2. The transition probability is given by an empirical method in the determination of the state transition probability determined by the way of transition of the state showing the maximum state output probability.
1. The voice recognition method described in 1 . 3. In the determination of the state transition probability determined by the way of the transition of the state showing the maximum state output probability, all the states of the total hidden Markov model are decomposed, and all the states are fixed while the inside which determines the output probability is fixed. giving initialized freely transition, again hidden by the speech data learned Markov model speech recognition method of claim 1, wherein the learning the transition probability. 4. A voice recognition device for recognizing a voice by using a hidden Markov model, comprising means for acoustically analyzing input voice to obtain a voice parameter, and each state of each hidden Markov model of the analyzed voice parameter. A means for calculating the output probability for, a language search using the language information such as a dictionary or grammar, the output probability of the hidden Markov model and its state, and means for obtaining the recognition candidate of the input speech and its likelihood. From the result of calculating the output probability for each state of the hidden Markov model, the means for obtaining the maximum state output probability and its state for the analyzed speech parameter, and the rejection for the determined state transition method .
And a means for determining a likelihood for rejecting a voice input based on the state transition probability, and a means for determining whether to output or reject a recognition result of the voice input based on a ratio of the two likelihoods. A voice recognition device characterized by the above. 5. The means for obtaining the likelihood for rejecting is characterized in that a transition probability is given by an empirical method, and the state transition probability is determined by the way of transition of the state showing the maximum state output probability. Item 4. The voice recognition device according to item 4 . 6. The means for obtaining the likelihood for rejecting decomposes all the states of the total hidden Markov model, and gives initialization so that all the states can freely transition while fixing the inside that determines the output probability, 5. The voice recognition device according to claim 4 , wherein the transition probability is learned from the voice data from which the hidden Markov model is learned again, and the state transition probability is determined by the way of transition of the state showing the maximum state output probability.