JPH0594198A - Method and device for recognizing voice - Google Patents

Abstract

PURPOSE:To improve the recognition accuracy of phonemes having a transient feature such as a plosive, a nasal sound by deriving separately a covariance matrix for representing the variance of a normal distribution for every state. CONSTITUTION:The device is provided with a microphone 1 for inputting voice information, an A/D converter 2 for executing analog/digital conversion (A/D conversion) of the voice information inputted from the microphone 1, and a CPU(Central Processing Unit) 3 for taking charge of control of each part. Also, a read-only memory(ROM) 4 stores an average value and a covariance of each standard pattern, and moreover, stores a program of a processing, and a random access memory(RAM) 5 is used as a memory for a work space, and each part is connected by a system bus 6. In such a state, at the time of extracting an analytic parameter from a sound signal, a dynamic feature parameter and a static feature parameter are both derived, and at the time of DP matching, pattern matching is executed, based on the respective output probability.

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 apparatus for recognizing phonemes, syllables, words, etc. from parameters obtained by analyzing speech.

[0002]

2. Description of the Related Art Conventionally, as a method for recognizing voice, H
Many means such as the MM method for recognizing speech by a statistical method are used. One of these methods is the stochastic DP method, which is a recognition method based on DP matching using a statistical distance measure. The Stochastic DP method is referred to by Nakagawa, "Recognition of English Consonants of Unspecified Speakers by Stochastic DP Method and Statistical Method", IEICE Transactions (D), J70-D, 1, p. p. 1
55-163 (Sho 62-01), the distance measure corresponds to the probability measure, and the transition probability is used instead of the path cost.

[0003]

However, in the above-mentioned conventional method, since one covariance matrix is used for one phoneme category, there is a drawback that a phoneme having a transient feature cannot be fully expressed. there were.

Further, since a static feature which is a vector quantity representing an absolute position of a voice spectrum is used as a voice parameter, there is a drawback that a phoneme having a transient feature is difficult to recognize.

Further, since this method basically uses the DP matching method, it is possible to change to the spotting algorithm by using an asymmetric DP path having the standard pattern as the basic axis. This method is described in Nakagawa, “Speech Recognition by Probabilistic Model”, edited by IEICE, pp. 87-89 (Sho 63-7). However, since the number of times the output probability is calculated is not constant, there is a drawback in that if the standard patterns have different lengths, they are likely to be recognized as short patterns.

[0006]

In order to solve the above-mentioned problems, the present invention uses a statistical distance measure based on the assumption of normal distribution as a distance measure of a spectrum, and speech by dynamic programming using a state transition probability. Provided is a recognition device, which is characterized in that a covariance matrix representing the variance of a normal distribution is obtained separately for each state.

In order to solve the above-mentioned problems, the present invention is characterized in that the speech recognition apparatus has an analysis means for analyzing input speech to obtain parameters, and the analysis uses both static features and dynamic features. The voice recognition device according to claim 1, wherein

In order to solve the above problems, the present invention provides a speech recognition method and apparatus by dynamic programming that uses a statistical distance measure based on the assumption of normal distribution as a distance measure of a spectrum and uses a state transition probability. A speech recognition method and apparatus, wherein a covariance matrix representing the variance of a normal distribution is obtained separately for each state.

In order to solve the above-mentioned problems, the present invention uses the parameters obtained by analyzing the input voice for the recognition of the voice, derives the static feature and the dynamic feature by the analysis, and recognizes the feature. It should be used together.

In order to solve the above problems, the present invention places the integration axis of the dynamic programming on the side of the standard pattern and normalizes it by the number of states.

[0011]

Preferred embodiments of the present invention will be described in detail below with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of the speech recognition apparatus of this embodiment. In the figure, 1 is a microphone for inputting voice information, 2 is an A / D converter that performs analog / digital conversion (A / D conversion) of voice information input from the microphone 1, and 3 is a CPU (Central Process).
g Unit) and controls the control of each of these parts. A read-only memory (ROM) 4 stores the average value and covariance of each standard pattern of voice, and also stores a processing program as shown in a flowchart described later. A random access memory (RAM) 5 is used as a workspace memory. Reference numeral 6 denotes a system bus, and the above-mentioned units are connected by this system bus.

Now, how to obtain the average value of the standard patterns stored in the ROM 4 will be described.

As a standard pattern learning method, first, LPC is used.
DP matching between the standard pattern and the input parameter sequence is performed using a voice parameter such as cepstrum. Here, as the initial value of the standard pattern, an arbitrary voice parameter sequence belonging to the class to be statistically selected is selected. If alignment is achieved by DP matching, each frame of the standard pattern is regarded as a state, and the sum of the input parameters corresponding to each state and the sum of squares are calculated for each state. The average value for each state is calculated from the sum of the input parameters. The average value is updated every input. When all data are input, the covariance matrix is calculated using the sum of squares. In addition, when alignment is taken, the number of times the DP path is selected for each state is counted, and the state transition probability is calculated at the time when all the data are input.

Since the alignment between the input parameter series and each state is obtained by DP matching, if there is one input data, each state is always calculated once. For this reason, if the data of about the square of the order is required for the calculation of the covariance matrix, for example, assuming that the parameter order is the 12th order, the covariance matrix for each state can be designed with 144 data.

The average value and the state transition probability of each state of the standard pattern thus obtained are stored in the ROM 4.

Next, a process of recognizing a voice using the standard pattern data will be described with reference to the flowchart of FIG. The process of the flowchart of FIG. 2 is a process after the voice information input from the microphone 1 is converted into a digital signal by the A / D converter 2.

The voice signal has a static feature parameter (eg LPC cepstrum) which is a vector quantity representing an absolute position of the voice spectrum and a dynamic feature parameter which is a vector quantity representing a temporal movement of the voice spectrum. (For example, delta cepstrum) (S-
1).

At S2, the initial value of DP matching is given by using the equations (1-1) to (1-3).

Q (−1, j) = Q (0, j) = − ∞ Formula (1-1) Q (i, 1) = logP (a _i | 1) Formula (1-2) Q (i, 0) = 0 Expression (1-3) (where a _i represents a vector) In S3, it is determined whether or not the DP calculation for all classes (words, phonemes, etc.) has been completed.

If it is determined in S3 that the processing has not ended, the processing proceeds to S4, and DP is calculated using the gradual addition equations of equations (2-1) to (2-3). Required DP value (D
(Cumulative distance of P) Q (i, j) is expressed by equations (2-1) to (2).
-3) maximum value.

Q (i−2, j−1) + 0.5logP (a _i−1 | j) + 0.5logP (a _i | j) + logP ₁ (j) Formula (2-1) Q (i−1, j) j-1) + logP (a i | j) + logP 2 (j) equation (2-2) Q (i-1 , j-2) + logP (a i | j-1) + logP (a i | j) + logP 3 (J) Formula (2-3) where P _i (j) is the transition probability to the state j.

Note that P (a _i | j) representing the output probability used in the equations (2-1) to (2-3) is
It can be expressed as the following formula (3).

P (a _i | j) = λP _CEP (a _i | j) + (1−λ) P _DCEP (a _i | j) (0 ≦ λ ≦ 1) (3) where P _CEP (a _i | j) is the output probability due to the static feature parameter, and P _DCEP (a _i | j) is the output probability due to the dynamic feature parameter.

That is, since P (a _i | j) is used in the equations (2-1) to (2-3), the parameter obtained based on this equation should be a combination of the static feature and the dynamic feature. You can

In equation (3), P _CEP (a _i | j)
And the output probabilities used as P _DCEP (a _i | j) are represented by the following equation (4).

(2π) ^{−d / 2} | Σj | ⁻¹ / ² · exp {− (a _i −μ _j ) ^t Σ _j ⁻¹ (a _i −μ _j )} (4) where d is the dimension number of the parameter, μ _j is the average value vector in the state j, Σ _j is the covariance matrix in the state j, and a _i is the vector.

The above calculation is repeated until it is determined in S3 that the DPs of all the classes have been completed. If it is determined that the DPs of all the classes have been calculated in S3, the process proceeds to S4.

In S4, in order to normalize the number of states, Q (i,
J) / J (where J is the number of model states) is calculated for each standard pattern, and the class to which the standard pattern giving the maximum belongs is time i
The recognition class (phoneme, word, etc.) in. (J is the number of states of the model) The above equations (1) to (4) are calculated by the CPU 3. The average value and covariance of each standard pattern is R
It is stored in OM4.

[0030]

As described above, the covariance is provided for each state, and by using the parameter showing the static feature and the parameter showing the dynamic feature together, transient features such as plosive sounds and nasal sounds can be obtained. The accuracy of recognizing the phoneme possessed is improved. Further, by normalizing with the number of states, the performance when spotting is improved.

[Brief description of drawings]

FIG. 1 is a flowchart showing a process of an embodiment of the present invention.

FIG. 2 is a block diagram showing a configuration of a voice recognition device of the present invention.

[Explanation of symbols]

 1 Microphone 2 A / D converter 3 CPU 4 ROM 5 RAM 6 System bus

Claims (6)

[Claims] 1. A speech recognition apparatus by dynamic programming using state transition probabilities, which uses a statistical distance measure based on the assumption of normal distribution as a spectral distance measure, and a covariance matrix representing the variance of a normal distribution. A voice recognition device characterized in that is obtained separately for each state. 2. The voice recognition device has an analysis unit for analyzing input voice to obtain a parameter, and the static feature and the dynamic feature are derived by the analysis and are used together during recognition. The voice recognition device according to claim 1. 3. The integration axis of the dynamic programming is placed on the side of the standard pattern and is normalized by the number of states.
The voice recognition device described in 1. 4. A speech recognition method by dynamic programming using state transition probabilities, wherein a statistical distance measure based on the assumption of normal distribution is used as a distance measure of the spectrum, and a covariance matrix representing the variance of the normal distribution. A method for recognizing speech, characterized in that is calculated separately for each state. 5. The voice recognition is performed by using a parameter obtained by analyzing an input voice, deriving a static feature and a dynamic feature by the analysis, and using them together for recognition. Speech recognition method described in. 6. An integration axis of the dynamic programming is placed on the side of a standard pattern and is normalized by the number of states.
Speech recognition method described in.