JPH0642153B2 - Voice recognizer - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a speech recognition apparatus, and more particularly to a speech recognition apparatus using a discrete relaxed Markov model (hereinafter referred to as HMM).

[Prior Art] FIG. 5 is a diagram showing the principle of a conventional speech recognition apparatus using an HMM. With reference to FIG. 5, the codebook creating means 1 finds a set of a plurality of representative vectors that most closely approximates them from a large number of speech feature vectors, and the obtained plurality of representative vectors are encoded by a coding means. Give to 3. The encoding means 3 encodes the speech feature vector using the given set of a plurality of representative vectors as a codebook, and supplies it to the HMM training means 4 and the HMM recognition means 5. The HMM training means 4 trains a discrete HMM by using a plurality of code time series obtained by encoding a plurality of voice feature vector time series as training data. On the other hand, the HMM recognizing means 5 recognizes the code time series obtained by encoding the voice feature vector time series as the recognition data by the HMM trained by the HMM training means, and outputs the recognition result.

[Problems to be Solved by the Invention] In the speech recognition apparatus shown in FIG. 5 described above, the speech feature vector is trained to have a high generation probability in the correct category, but the probability in the wrong category is increased. There is no training to reduce. Therefore, there is a problem that high voice recognition performance cannot be obtained.

Therefore, a main object of the present invention is to provide a speech recognition apparatus using a discrete HMM capable of obtaining high speech recognition performance.

[Means for Solving the Problem] FIG. 1 is a diagram showing the principle of the present invention, and is a codebook creating means 1 for obtaining a set of a plurality of representative vectors that most approximates a large number of speech feature vectors. , A category name is given to each representative vector forming the codebook so that the number of coincidences between the representative vector category and the speech feature vector category used when encoding a plurality of speech feature vectors increases. In the codebook learning means 2 for sequentially updating the representative vector, and using a set of a plurality of representative vectors as a codebook, the code number of the representative vector having the closest Euclidean distance to the input speech feature vector is code time series. And a plurality of code time series obtained by coding a plurality of voice feature vector time series, the generation probability becomes maximum. Such transition probabilities and output probabilities are obtained, and discrete HM is used as training data.
The training means 4 for training M and the code time series obtained by coding the voice feature vector time series are input, and the transition probability and the output probability are combined to generate the input code time series. The recognition unit 5 calculates and outputs the voice with the highest generation probability.

[Operation] The speech recognition apparatus according to the present invention obtains a set of a plurality of representative vectors that most approximate these from a large number of speech feature vectors, assigns a category name to each representative vector, and determines a plurality of speech feature vectors. The representative vector is sequentially updated so that the number of coincidences between the category of the representative vector used for encoding and the category of the speech feature vector increases, and the set of these plurality of representative vectors is used as a codebook for speech. A feature vector is encoded, a discrete HMM is trained using a plurality of code time series obtained by encoding as training data, and recognition is performed by the trained HMM.

Embodiment of the Invention FIG. 2 is a schematic block diagram of an embodiment of the present invention.
In this embodiment, the case of performing phoneme recognition of Japanese 23 phonemes will be described. The phoneme data 11 consists of 100 phoneme samples for each phoneme category.
One phoneme sample consists of a time series of phoneme feature vectors. The phonological feature vector is, for example, a 16-dimensional power spectrum. The K-means clustering means 12
All phoneme feature vectors of the phoneme sample for learning,
Clustering is performed for each phoneme category using the K-means clustering method, and 10 representative vectors are obtained for each phoneme category. Then, a total of 230 representative vectors obtained for all phoneme categories is set as the codebook 13.

Each phoneme category name is given to each representative vector of the codebook 13. The LVQ learning unit 14 uses the learning vector quantization method (hereinafter, referred to as LVQ) to determine the number of coincidences between the category of each representative vector and the category of the phoneme feature vector used when encoding the phoneme feature vector. The representative vector is sequentially updated so as to increase.

FIG. 3 is an LV by the LVQ learning means 14 shown in FIG.
It is a flowchart which shows a Q2 learning algorithm. In step (abbreviated as SP in the drawing) SP1, a representative vector m _i having the smallest Euclidean distance to the phoneme feature vector x and a representative vector m _j having the smallest Euclidean distance in categories other than the category of the representative vector are obtained. . At step SP2, it is determined whether or not the representative vector is updated. The condition is
That is, it does not match the category to which m _i belongs, and the category to which m _j belongs matches the category to which x belongs. In step SP3, the representative vectors m _i and m _j are updated only when the condition in step SP2 is satisfied.

The update is performed by the following formula.

m _i ′ = m _i −a (t) (x−m _i ) m _j ′ = m _j + a (t) (x−m _j ), where m _i ′ and m _j ′ represent the updated representative vectors. ,
a (t) is a function that monotonically decreases with time (a
(T)>0>).

The LVQ learning means 14 shown in FIG. 2 performs the above-mentioned operation for all phoneme feature vectors for learning, and repeats it for an appropriate number of times. The encoding means 15 encodes all phoneme samples (phoneme feature vector time series) to obtain phoneme sample code time series. This encoding is
This is performed by outputting the code number of the representative vector having the closest Euclidean distance to the input feature vector.

The HMM training means 16 inputs the code time series of the phoneme samples for learning and trains each phoneme model 17 using the Baum-Welch algorithm.

FIG. 4 is a diagram showing the structure of a phoneme model. Referring to FIG. 4, s indicates a state, a indicates a transition probability, and b indicates an output probability. For example, in the state s ₁ , the probability of staying in the state s ₁ is a ₁₁ , and the probability of transiting to the state s ₂ is a ₁₂ . Also, the probability of outputting the code k when a transition from or state s ₁ time remained state s ₁ to the state s ₂ is b _{1, k,} 1 subscript is a transition that is output from the state s ₁ Is shown. HM
The M training unit 16 obtains, for each phoneme, a transition probability a and an output probability b that maximize the generation probability of each input phoneme sample code time series, and outputs them as a phoneme model.

The HMM recognizing means 18 receives the code time series of the phoneme sample to be recognized as input, uses the forward pass algorithm, and multiplies the transition probability a and the output probability b to calculate the probability of generating the input code time series as a whole phoneme model. Calculate about.
Then, the phoneme model with the highest generation probability value is obtained,
Output as a phoneme recognition result.

[Effects of the Invention] As described above, according to the present invention, in a voice recognition device using a discrete HMM, a category name is assigned to each representative vector forming the codebook, and a plurality of voice feature vectors are provided. Since the representative vectors are updated sequentially so that the number of coincidences between the category of the representative vector and the category of the speech feature vector used when encoding the Can be created, and the voice recognition performance can be improved.

[Brief description of drawings]

FIG. 1 is a block diagram for explaining the principle of the present invention. FIG. 2 is a schematic block diagram of an embodiment of the present invention. FIG. 3 is a flow chart showing the LVQ2 learning algorithm shown in FIG. FIG. 4 is a diagram showing an HMM phoneme model. FIG. 5 is a diagram for explaining the principle of a conventional HMM voice recognition device. In the figure, 1 is a codebook creating means, 2 is a codebook learning means, 3 is an encoding means, 4 is an HMM training means, and 5 is an HMM.
Recognition means, 11 is phoneme data, 12 is K-means clustering means, 13 is a codebook, 14 is LVQ learning means, 1
Reference numeral 5 is an encoding means, 16 is an HMM training means, 17 is a phoneme model, and 18 is an HMM recognition means.

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Eric McDermott, 5 Seiraya, Seika-cho, Soraku-gun, Kyoto Prefecture, Mihiratani, A-R Co., Ltd. (56) References Acoustics Society of Japan October of the first year 1-1-20 P. 39-40

Claims (1)

[Claims] 1. A codebook creating means for finding a set of a plurality of representative vectors that most closely approximate these from a large number of speech feature vectors, and a codebook for the set of a plurality of representative vectors found by the codebook creating means. As coding means for outputting the code number of the representative vector having the closest Euclidean distance to the input speech feature vector as a code time series, and obtained by encoding a plurality of speech feature vector time series by the coding means. For a plurality of code time series, a transition probability and an output probability that maximize the generation probability are obtained, and training means for training a discrete Hidden Markov Model as training data, a speech feature vector time series by the coding means. The code time series obtained by encoding is input, the transition probabilities and output probabilities are combined, and the input code time system is input. Recognizing means for calculating the probability of generating, and obtaining and outputting the voice with the highest generation probability, and a category name is given to each representative vector constituting the codebook created by the codebook creating means, Speech recognition provided with a codebook learning means for sequentially updating the representative vectors so that the number of coincidences between the category of the representative vector used when encoding apparatus.