JPH0981182A - Learning device for hidden markov model(hmm) and voice recognition device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reduce the time and the cost required to change the vocabulary of registered words by generating pseudo word learning data that are produced based on uniform random numbers and relearning garbage hidden HMMs. SOLUTION: While changing the vocabulary of registered words, an ME learning method, in which a garbage HMM 12 is learned using pseudo word learning data as the learning method of the HMM 12 that does not require the collection and the manipulation of new voice samples, is employed. In this case, the pseudo word learning data are generated based on a prescribed word HMM 11 which is a phoneme HMM (a CD phoneme HMM) considering the front and behind phoneme environment and uniform random numbers generated by a digital electronic computer. By generating pseudo word learning data which are generated based on uniform random numbers and relearning the HMM 12 based on the data above, no work is required for the collection and the manipulation of new learning data voice samples.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an HMM learning device for learning a hidden Markov model (hereinafter referred to as HMM) for speech recognition, and speech recognition using the HMM learned by the learning device. Voice recognition device.

[0002]

2. Description of the Related Art In a conventional continuous speech recognition apparatus, an HMM is used to perform speech recognition based on acoustic feature parameters extracted from an input uttered speech, and the result is output.

In the extraction (also referred to as spotting) of registered words using the HMM by the voice recognition unit, the garbage HMM used when detecting unregistered words has a great influence on the spotting performance. Conventionally, garbage H
A learning method based on an error minimization criterion (hereinafter referred to as ME learning method) is used to learn MM, and its effectiveness is described in, for example, the conventional document 1 “Komori et.
al. , "Minimum error classi
fiction training for HMM-
based keyword spotting ", P
rc. ICSLP 92, Vol. I, pp. 9-1
2, 1992 "and the conventional document 2" Torre et.
al. , "Discriminative train
ingof garbage model for n
on-vocabulary utterance re
injection ”, Proc. ICSLP 94, Vo
l. I, pp. 475-478, 1994 ".

[0004]

However, in these conventional methods, since a large amount of voice samples are used for learning, it is necessary to collect and cut voice samples each time the vocabulary of registered words is changed. , It was impossible in principle to change the vocabulary of registered words quickly.

An object of the present invention is to solve the above problems and to learn an HMM by changing the vocabulary of registered words more easily and quickly than in the conventional example, and an HMM learning device and the learning device. Another object of the present invention is to provide a voice recognition device for recognizing a voice using the HMM learned by.

[0006]

A hidden Markov model learning apparatus according to a first aspect of the present invention is a random number generating means for generating a multidimensional uniform random number, and a multidimensional one generated by the random number generating means. -Like random numbers are converted into a plurality of Gaussian random numbers that follow the multidimensional Gaussian distribution of a given word hidden Markov model for recognizing preregistered registered words, and the converted plurality of Gaussian random numbers are a plurality of feature parameters. Data generation means for outputting as pseudo word learning data, pseudo word learning data output from the data generating means, based on the word hidden Markov model,
Learn a plurality of parameters of the Garbage Hidden Markov Model by updating a plurality of parameters of the Garbage Hidden Markov Model for detecting unregistered unregistered words so that the function value of the predetermined cost function is minimized. And a learning means.

A hidden Markov model learning apparatus according to a second aspect is the hidden Markov model learning apparatus according to the first aspect, wherein the cost function includes the word learning data and the word hidden for a word to be recognized. Spoken, calculated based on a score for speech recognition calculated based on the Markov model and a score for speech recognition calculated based on the word learning data and the garbage hidden Markov model. An index value indicating the possibility of an error in which a word is not recognized, and a recognition target word, for speech recognition calculated based on the word learning data excluding the recognition target word and the word hidden Markov model A score and a score for speech recognition calculated based on the word learning data and the garbage hidden Markov model It was calculated Zui, characterized in that it is a function calculated by word to be uttered is added to, and the index value indicating the likelihood of errors appearing in the recognition result.

Further, in the hidden Markov model learning apparatus according to claim 3, in the hidden Markov model learning apparatus according to claim 2, the score indicating the likelihood for the speech recognition is calculated by the Viterbi decoding method. It is characterized by the score.

The speech recognition apparatus according to the present invention is a registered word registered in advance based on a hidden Markov model learning apparatus according to claim 1, 2 or 3 and a voice signal of an input uttered voice sentence. To recognize the word Hidden Markov model, and using a garbage Hidden Markov model for detecting unregistered words that are learned by the learning device of the hidden Markov model and not registered in advance, and output a speech recognition result. It is characterized in that a voice recognition means is provided.

[0010]

DETAILED DESCRIPTION OF THE INVENTION Embodiments of the present invention will be described below with reference to the drawings. In the present invention, in order to enable the vocabulary set of registered words to be changed quickly and easily, it is not necessary to collect or process a new voice sample when changing the vocabulary set of registered words. An ME learning method for learning a garbage HMM using learning data (hereinafter, word learning data) is used. Here, the pseudo word learning data is generated based on a predetermined word HMM11 which is a phoneme HMM (hereinafter, a CD phoneme HMM) considering the phoneme environment before and after and a uniform random number generated by a digital electronic computer. .

FIG. 1 is a block diagram of a speech recognition apparatus according to an embodiment of the present invention. The registered word spotting algorithm used in the word matching unit 4 is a one-pass Viterbi decoding method (One-pass Viterbi decode).
coding). As the acoustic model, as shown in FIG. 9, an HMM for each registered word formed by concatenation of CD phoneme HMMs, a silent HMM in one state, and a garbage HMM are used. Here, the HMM and the silent HMM relating to each registered word for recognizing a plurality of registered words registered in advance are stored in the memory of the word HMM 11 in FIG. 1 while detecting an unregistered word not registered in advance. The garbage HMM is the garbage HMM shown in Fig. 1.
It is stored in 12 memories. These HMMs 11, 12
The memory is composed of, for example, a hard disk memory.

The word HMM 11 has a connection relation between states shown in FIG. 8A and an information structure shown in FIG. 8B. The word HMM11 is, as shown in FIG.
Multiple n states are represented by cascaded state transitions,
It has a self-loop in each state. And the word HMM11
As shown in FIG. 8 (b), each state of includes a self-loop probability, state transition probability, and output distribution data, and the output distribution data includes a multidimensional Gaussian distribution number, a mixture weight, a dimension number, Contains the mean and variance values for each dimension. Here, the multidimensional Gaussian distribution means, for example, 16-dimensional LPC cepstrum, 16-dimensional Δ cepstrum, logarithmic power, Δ
It is a Gaussian distribution regarding a 34-dimensional feature parameter including logarithmic power.

FIG. 2 shows a garbage HMM learning process executed by the HMM learning unit 20 shown in FIG. CD phoneme H
The word HMM11 including MM and silence HMM and the garbage HMM12 are known Baum-Welch (Baum-We).
lch) algorithm, the parameters are learned in advance and the initial parameters are set, and the garbage HMM1
Only 2 is re-learned by the ME learning method by the process of FIG.

As shown in FIG. 2, first, in step S1, a word learning data creating process for creating pseudo word learning data is executed. Here, the word learning data is the word HMM11 for each word to be recognized.
It is generated based on the information that each word HMM in has. Specifically, for each word, a uniform pseudo-random number generated by a digital computer is converted into a random number that follows a multidimensional Gaussian distribution determined by the state transition rule of the word HMM and the mixture weight distribution of each state. It is realized by repeating the output procedure until the final state of each word HMM is reached.

Next, in step S2, the function value of the cost function defined so as to correspond to the index value of the recognition error occurrence probability, which will be described later in detail, based on the created word learning data is the minimum value (actual value). To obtain a new garbage HMM, each parameter of the garbage HMM 12 is sequentially updated so that it becomes a local minimum value. Actually, a plurality of word learning data sets are prepared, and learning is advanced so as to minimize the average value of costs for each word learning data set.

The cost function used in step S2 is calculated based on the difference in the score of Viterbi using the word learning data, the word HMM11, and the garbage HMM12 which is an HMM for detecting unregistered words. , Is defined as an index value of recognition error occurrence probability. There are two types of recognition errors: an uttered word is not recognized (that is, a word drop error) and an unspoken word appears in the recognition result (that is, a word-swelling error). The index value is calculated by the method described later in detail. The cost function C is the following equation 1
It is represented by

[0017]

[Equation 1]

Here, E (Δ) represents a sigmoid function expressed by the following equation 2.

[Equation 2] E (Δ) = 1 / (1 + exp (−αΔ))

Further, P (S, x) is a function expressed by the following equation 3, and indicates the maximum value of the Viterbi score V (s, x) of the HMMs with respect to the word learning data x.

(Equation 3)

Further, ghω is a function expressed by the following equation 4, and is an argument for maximizing the Viterbi score V (ω, w).

(Equation 4)

Further, it is as follows. W: set of word learning data, | W |: number of words in set of word learning data, Ω: set of word HMM11, | Ω |: number of words in set of word HMM11, γ: set of garbage HMM12, V (s , X): Viterbi score of the HMMs for the word learning data x, h: each element of the set W of word learning data (that is, each word learning data), each element of the set Ω of the corresponding word HMM11 (that is, the word) A bijective shot to each HMM in HMM11).

In the cost minimization process in step S2, each parameter θ (that is, the average, the variance, and the mixing weight) of the garbage HMM 12 is sequentially updated by the following equation 5 until the cost function value converges. .

[0023]

[Equation 5] θ ⁽ⁱ⁾ = θ ^(i-1) −β [∂C / ∂θ] (θ = θ ^(i-1) )

Here, θ ⁽ⁱ⁾ is a parameter obtained by the i-th update, and β is a learning constant, which takes a value of 0.1 to 0.5, for example. In addition, [∂C / ∂θ] (θ = θ ^(i-1) ) of the second term on the right side of Equation 5 is θ = θ ^(i-1)
It is [∂C / ∂θ] at the time.

FIG. 3 is a flow chart showing the word learning data generation process (step S1) of FIG. As shown in FIG. 3, the word number j is set to 1 in step S11, and in step S12, the characteristic parameter string generation process for the word of the word number j (hereinafter, referred to as word #j), which will be described later in detail, is executed. . Next, the phoneme duration of the characteristic parameter string generated in step S13 is checked, and it is determined in step S14 whether the phoneme duration is normal. As a concrete judgment criterion at this time, 20 ms or less was judged to be normal in the case of vowels, and 10 ms or less was judged to be normal in the case of consonants. If the phoneme duration is normal, the characteristic parameter string generated in step S15 is output to the working memory 21 and temporarily stored, and the word number j is set to 1 in step S16.
It increments by one and proceeds to step S17. In step S17, it is determined whether or not the generation of the word learning data has been completed for all the words. If not, the process returns to step S12 to repeat the above process, and if the generation is completed, the word learning data generation is completed. The process ends. If it is determined in step S14 that the phoneme duration is not normal, the generated characteristic parameter sequence is removed without being output, and another uniform random number is generated to generate another characteristic parameter sequence. Returning to step S12, the above-mentioned processing is repeated.

FIG. 4 is a flowchart showing the characteristic parameter string generation process (step S12) for the word #j in FIG. As shown in FIG. 4, first, in step S21, the state number i is set to 1, and in step S22, a uniform random number generated by the digital electronic computer that is the HMM learning unit 20 (the uniform random number is 0 to 1). Between the values up to) and the state of state number i (hereinafter referred to as state #i).
The multidimensional Gaussian distribution number j of the word #j in the word HMM11 is determined according to the mixture weight distribution of. Ie the word #
Since the sum of the mixture distributions of a plurality of multidimensional Gaussian distributions of j is 1, it is determined whether the value of the generated uniform random number corresponds to the cumulative addition value of the mixture weights of the multidimensional Gaussian distributions. By doing so, the multidimensional Gaussian distribution number j of the word #j is determined. For example, the cumulative addition value of the mixing weights in the multidimensional Gaussian distribution number # 2 is a value obtained by adding the mixing weights in the multidimensional Gaussian distribution number # 2 and the mixing weights in the multidimensional Gaussian distribution number # 1 and is generated. When the uniform random number exceeds the mixing weight in the multidimensional Gaussian distribution number # 1 and is equal to or less than the cumulative addition value of the mixing weights in the multidimensional Gaussian distribution number # 2, j = 2 is determined.

Next, in step S23, the multidimensional uniform random number generated by the digital computer is converted into the multidimensional Gaussian distribution of the Gaussian distribution number j of the word #j in the word HMM11 (hereinafter, the multidimensional Gaussian distribution #j). It is converted into a plurality of Gaussian random numbers (also referred to as normal random numbers) according to the above), and the result is output to the working memory 21 as a characteristic parameter string. Here, the Gaussian random number according to the multidimensional Gaussian distribution #j is a Gaussian random number having the same mean, variance, and shape of the Gaussian distribution.

Further, in step S24, the presence or absence of state transition is determined based on the uniform random number generated by the digital computer and the transition probability of state # 1 of word #j in word HMM11. That is, when the generated uniform random number is less than or equal to the transition probability, it is determined that the state transition occurs, and when the generated uniform random number exceeds the transition probability, it is determined that the state transition does not occur. Next, in step S25, it is determined whether or not the state transition occurs. If the state transition does not occur, it is determined as a self-loop, and another uniform random number is generated to generate another characteristic parameter sequence in step S2.
Return to 2. On the other hand, when it is determined in step S25 that the state transition occurs, the state number i is incremented by 1 in step S26, and the state #i is changed to the HM in step S27.
It is determined whether or not it is the final state of M. If it is not the final state, the process returns to step S22, the above-described processing is repeated for the next state, and if it is the final state, the process returns to the main routine of FIG.

FIG. 5 is a flow chart showing a cost function calculation process which is a subroutine executed in step S2 of FIG. As shown in FIG. 5, first, in a step S31, a word drop error occurrence probability index value calculation process, which will be described later, is executed, in a step S32, a word spelling error occurrence probability index value calculation process is executed, and then in a step S
In 33, the value calculated in step S31 and stored in the calculation buffer Buff1 is added to the value calculated in step S32 and stored in the calculation buffer Buff2, and the addition result is set as the cost function value C.

FIG. 6 is a flow chart showing the word drop error occurrence possibility index value calculation processing of FIG. In this process, for each recognition target word, a Viterbi score is calculated based on the pseudo word learning data and the HMM of the word in the word HMM11, and based on the pseudo word learning data and the garbage HMM12. The Viterbi score is calculated, the result obtained by subtracting the Viterbi score by the HMM of the word from the Viterbi score by the garbage HMM12 is smoothed by the sigmoid function, and the sum of the smoothed values is divided by the number of recognition target words to obtain the normal value. It is converted into the word omission error possibility index value.

As shown in FIG. 6, 0 is set in the calculation buffer Buff1 in step S41, and the calculation buffer Buff1 is set in step S42.
After the word number j is set to 1 in step S43, in step S43, the Viterbi score by the HMM of the word is subtracted from the Viterbi score by the garbage HMM12, which is the right side of Σ in the first term on the right side of the equation 1. Then, the calculated value is added to the value of the calculation buffer Buff1, and the addition result is updated as the value of the calculation buffer Buff1. Then, in step S44, the word number j is set to 1
It is determined whether or not the process of step S43 has been completed for all the words in step S45, and if not completed, the process returns to step S43 to repeat the above process, and if completed, the step of S
Proceed to 46. In step S46, the calculation buffer Buf
The value of f1 is divided by the number of words in the set of word learning data, and the result of the division is stored in the calculation buffer Buff1. Finally, in step S47, the value of the calculation buffer Buff1 is output to the working memory 21 and stored as the word loss error occurrence possibility index value corresponding to the first term on the right side of the equation 1.

FIG. 7 is a flow chart showing the process of calculating the word omission error occurrence possibility index value of FIG. In this process, for each recognition target word, a Viterbi score is calculated based on all the word learning data excluding the word and the HMM of the word in the word HMM11, and the maximum Viterbi score among these calculated words is calculated. The word learning data ghω and the score y thereof are stored, and the Viterbi score z is calculated based on the word learning data ghω and the garbage HMM12, and the garbage HMM1 is calculated from the score y.
A value obtained by subtracting the Viterbi score z calculated based on 2 is smoothed by a sigmoid function, the sum of the smoothed values is divided by the number of words to be recognized, and then normalized, and a word spring error probability index The value.

As shown in FIG. 7, 0 is set in the calculation buffer Buff2 in step S51, and the calculation buffer Buff2 is set in step S52.
After the word number j is set to 1 in step S53, in step S53, based on the HMM of the word #j in the word HMM11 (hereinafter referred to as the word HMM # j) and each word learning data excluding the word #j. The score of each Viterbi is calculated, and the word learning data ghω which gives the maximum value among these calculated scores is selected and stored in the working memory 21.
Next, in step S54, the Viterbi score z is calculated based on the word learning data ghω that gives the maximum value and the garbage HMM12, and the maximum Viterbi score, which is the part on the right side of Σ in the second term on the right side of Expression 1, is calculated. Garbage HM from y
A value obtained by subtracting the Viterbi score z calculated based on M12 is calculated, the calculated value is added to the value in the calculation buffer Buff2, and the addition result is calculated in the calculation buffer Buff2.
Update as the value of. Then, in step S54, the word number j is incremented by 1, and in step S55, it is determined whether or not the processes of steps S53 and S54 have been completed for all the words. If not completed, the process returns to step S53. The above process is repeated, and when the process is completed, the process proceeds to step S57. In step S57, the value of the calculation buffer Buff2 is divided by the number of words in the set of word HMM11, and the result of the division is calculated in the calculation buffer Buf.
Store in f2. Finally, in step S58, the value of the calculation buffer Buff2 is output to and stored in the working memory 21 as a word-shedding-error occurrence possibility index value corresponding to the second term on the right side of Expression 1.

Next, a speech recognition apparatus for word recognition that performs speech recognition using the garbage HMM 12 retrained by the above method and the word HMM 11 will be described with reference to FIG.

In FIG. 1, the HMM learning unit 20 relearns the garbage HMM 12 based on the pseudo word learning data and the word HMM 11, and stores it in the memory of the garbage HMM 12. On the other hand, the uttered voice of the speaker is input to the microphone 1 and converted into a voice signal, and then the feature extraction unit 2
Is input to The feature extracting unit 2 performs, for example, LPC analysis after A / D converting the input voice signal, and performs logarithmic power, 16th-order cepstrum coefficient, Δlogarithmic power, and 16 logarithmic power.
A 34-dimensional feature parameter including the next Δ cepstrum coefficient is extracted. The time series of the extracted characteristic parameters is input to the word matching unit 4 via the buffer memory 3.

The word collating unit 4 uses the known one-pass Viterbi decoding method based on the time-series data of the characteristic parameters stored in the buffer memory 3 to identify the word HMM 11 for recognizing the registered word and the word HMM 11 for recognizing the registered word. Using the garbage HMM12 for detecting a registered word, the Viterbi score for the data in the word matching section is calculated, and the word corresponding to the maximum Viterbi score is output as a recognized word string.

In the above embodiment, the word matching unit 4 and the HMM learning unit 20 are composed of, for example, a digital electronic computer.

[0038]

EXAMPLES The inventor of the present invention has found that the HMM learning unit 20 of the present embodiment.
The following experiment was conducted to confirm the effectiveness of.
The experimental conditions are shown in Table 1. As a CD phoneme HMM,
A 200-state speaker-independent HM network obtained by synthesizing a plurality of speaker-adapted hidden Markov networks (hereinafter referred to as HM networks) was used. As a target vocabulary for spotting,
A spontaneous vocal dialogue corpus owned by the applicant of the present invention whose task is travel planning such as hotel reservation (conventional document 3 “Morimoto et al.,“ A spe
ech and language database
for speech translation
research ”, Proc. ICSLP 94, Vo
l. IV, pp. 1791-1794, 1994 ". 20 words were selected from the above.

[0039]

[Table 1] Experimental conditions ─────────────────────────────────── Acoustic analysis conditions Sampling frequency = 12 kHz Number of bits = 16 bits Pre-emphasis = 1-0.97z ^-1 Hamming window = 20 ms Frame shift = 5 ms Feature parameter = 16-dimensional LPC cepstrum + 16-dimensional Δ cepstrum + power + Δ power ───────── ─────────────────────────── HMM Topology Word HMM: 3 states or 4 states, 5 mixed silence HMM: 1 state, 10 mixed garbage HMM ;: 1 state, 20 mixed ───────────────────────────────────

In the learning process by the ME learning method, it is important to start learning from the best possible initial model. In this experiment, an initial garbage HMM was created by synthesizing multiple speaker-specific models. In this method, a plurality of HMMs are created in order to secure the resolution for each of acoustic characteristics and speaker characteristics, and then they are combined into one HMM having a desired mixture number.

As word learning data, 20 sets of pseudo word learning data of all vocabularies were generated. In learning, learning data corresponding to unregistered words is also required. If statistical language data about unregistered words is available, create a language model that universally represents unregistered words, and apply the above-mentioned data generation method on it to create words for unregistered words. Learning data can be generated. In this experiment, as an application example of this method when these language data cannot be used, a method of selecting and using a substitute for the word learning data regarding unregistered words from the pseudo word learning data is used. It was That is, in the calculation of the word-swelling error probability index value for each registered word in Equation 1, the registered word in the pseudo word learning data excluding the registered word is substituted for the word learning data regarding the unregistered word. HMM
The one that gave the maximum Viterbi score for was used.
As a result, the garbage HMM 12 is learned so as to give each word learning data a score lower than that of the correct word HMM and higher than that of the incorrect word HMM.

Then, the created initial garbage HMM was re-learned by the above-mentioned learning method. In this experiment, the parameters of the average and the mixture weight were updated. The number of repeated calculations until the cost function value reached convergence was 20.

Next, the spotting experiment and the result will be described. Using the re-learned garbage HMM, a speaker open registration word spotting experiment was conducted for one speaker for each gender. As test data, 4 dialogues were selected from the above-mentioned spontaneous speech dialogue corpus. The total number of utterances is 6
It was 0, and the total number of appearances of registered words was 22. FIG.
As shown in FIG. 0, by using the ME retrained garbage HMM12 in the method of the present invention, the initial garbage HM
Compared with the case where M is used, the word error release rate characteristic with respect to the drop rate of registered words is improved. From this result, it is understood that the learning method of the present invention is effective in improving the spotting performance even when the word learning data regarding the substitute unregistered word is used.

As described above, according to the present embodiment, since the pseudo word learning data generated based on the uniform random number is generated and the garbage HMM12 is retrained based on the pseudo word learning data, the estimation target Since it is unnecessary to collect and process a new voice sample for learning data for a word, the time and cost required for changing the vocabulary set of registered words are significantly reduced. Therefore, it is possible to change the vocabulary set of registered words and re-learn the garbage HMM 12 more easily and quickly than in the conventional example. Further, when the voice recognition is performed using the re-learned garbage HMM 12, the voice recognition can be performed at a voice recognition rate almost equal to that of the conventional example.

[0045]

As described above in detail, according to the hidden Markov model learning apparatus of the present invention, the random number generating means for generating a multidimensional uniform random number and the multidimensional uniform generated by the random number generating means. Converts a random number into a plurality of Gaussian random numbers that follow a multidimensional Gaussian distribution of a given word hidden Markov model for recognizing preregistered registered words, and converts the plurality of converted Gaussian random numbers into a plurality of characteristic parameter pseudo-parameters. Based on the data generation means for outputting as typical word learning data, the pseudo word learning data output from the data generation means, and the word hidden Markov model so that the function value of the predetermined cost function is minimized. , By updating multiple parameters of the Garbage Hidden Markov Model for detecting unregistered words that are not registered in advance. And a learning means for learning a plurality of parameters of Dell. Therefore, pseudo word learning data generated based on uniform random numbers is generated, and the Garbage Hidden Markov Model is retrained based on the pseudo word learning data. Therefore, a new learning data voice sample is collected for the word to be estimated. Since no work such as or modification is required, the time and cost required for changing the vocabulary of registered words are significantly reduced.
Therefore, it is possible to change the vocabulary of the registered word and re-learn the Garbage Hidden Markov Model easily and quickly as compared with the conventional example.

Further, according to the speech recognition apparatus of the present invention, the hidden Markov model learning apparatus and a word for recognizing a registered word registered in advance based on the input voice signal of the uttered voice sentence. A speech recognition unit is provided which performs speech recognition using a hidden Markov model and a garbage hidden Markov model for detecting an unregistered word that is learned by the learning apparatus for the hidden Markov model and is not registered in advance, and outputs a speech recognition result. Therefore, by using the Garbage Hidden Markov Model that is relearned easily and quickly as compared with the conventional example, it is possible to perform voice recognition with a speech recognition rate almost equal to that of the conventional example.

[Brief description of drawings]

FIG. 1 is a block diagram of a voice recognition device according to an embodiment of the present invention.

FIG. 2 is a flowchart showing a garbage HMM learning process executed by the HMM learning unit 20 of FIG.

FIG. 3 is a flowchart showing a word learning data generation process which is a subroutine of FIG.

FIG. 4 is a flowchart showing a characteristic parameter sequence generation process which is a subroutine of FIG.

5 is a flowchart showing a cost function calculation process which is a subroutine executed in step S2 of FIG.

FIG. 6 is a flowchart showing a word loss error occurrence possibility index value calculation process which is a subroutine of FIG.

FIG. 7 is a flowchart showing a word-seeking error possibility index value calculation process which is a subroutine of FIG.

FIG. 8 is a diagram showing a structure of a word HMM, including (a)
[Fig. 3] is a state transition diagram showing a connection relation between states in the word HMM, and (b) is a diagram showing an information structure of the word HMM.

9 is a state transition diagram showing a spotting language model used in the speech recognition apparatus of FIG.

FIG. 10 is a graph showing a spring-out rate with respect to a drop-out rate of registered words, which is obtained in an experiment of the speech recognition apparatus of FIG.

[Explanation of Codes] 1 ... Microphone, 2 ... Feature extraction unit, 3 ... Buffer memory, 4 ... Word matching unit, 11 ... Word HMM, 12 ... Garbage HMM, 20 ... HMM learning unit, 21 ... Working memory.

Claims (4)

[Claims] 1. A random number generating means for generating a multidimensional uniform random number, and a predetermined word hidden Markov model for recognizing a registered word registered in advance for the multidimensional uniform random number generated by the random number generating means. Data generating means for converting the plurality of Gaussian random numbers according to the multidimensional Gaussian distribution of, and outputting the converted plurality of Gaussian random numbers as pseudo word learning data which are a plurality of feature parameters, and the data generating means. Based on the pseudo-word learning data and the word hidden Markov model, a plurality of garbage hidden Markov models for detecting unregistered words that are not registered in advance so that the function value of the predetermined cost function is minimized. A learning means for learning a plurality of parameters of the Garbage Hidden Markov Model by updating the parameters. Learning device of Re Markov model. 2. The cost function is a score for speech recognition, which is calculated based on the word learning data and the word hidden Markov model for a recognition target word, the word learning data and the garbage hidden Markov model. An index value that indicates the likelihood of an error in which a spoken word is not recognized, which is calculated based on the score for speech recognition calculated based on the model, and a recognition target word for the recognition target word. Score for speech recognition calculated based on the word learning data and the word hidden Markov model excluding, and score for speech recognition calculated based on the word learning data and the garbage hidden Markov model And an index value that indicates the likelihood of an error in which unvocalized words appear in the recognition result, calculated based on Learning apparatus Hidden Markov Models according to claim 1, characterized in that the function calculated by. 3. The hidden Markov model learning apparatus according to claim 2, wherein the score indicating the likelihood for speech recognition is a score calculated by a Viterbi decoding method. 4. The hidden Markov model learning device according to claim 1, 2 or 3, and a word hidden Markov model for recognizing a registered word registered in advance based on a voice signal of an input utterance voice sentence. And a voice recognition means for performing voice recognition using the Garbage Hidden Markov Model for detecting unregistered words that are learned by the hidden Markov model learning device and are not registered in advance, and output a voice recognition result. And a voice recognition device.