JPH0887292A - Word voice recognition device - Google Patents

Abstract

PURPOSE: To provide a word voice recognition device in which similar words pronounced by an unspecified speaker are quickly recognized with a high precision. CONSTITUTION: The device is provided with a voice input means 11 which detects uttered voice and inputs it as sound signals, an acoustic analysis means 12 which converts the signals into time series of a feature parameter based on the waveforms of the inputted sound signals, a feature point extracting means 13 which detects a feature time point at which the norm of the amount of spectrum changes becomes a maximum or a minimum from the time series of the feature parameter and extracts a feature point time series, a word standard pattern storage means 14 which stores the word standard patterns beforehand generated from word voice learning samples, a word collating means 15 which calculates the degree of similarity between the extracted feature point time series and the word standard patterns and a discriminating means 16 which outputs a recognition result from the degree of similarity of word.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a word voice recognition device, and particularly to a word voice recognition system capable of improving the recognition ability for similar words when recognizing the word voice of an unspecified speaker without deteriorating the recognition ability. Regarding the device.

[0002]

2. Description of the Related Art In word speech recognition of an unspecified speaker, how to absorb spectrum fluctuation and time length fluctuation due to individual differences is a major issue. In response to this problem, spectral fluctuations are standard patterns and time length fluctuations are DP.
Matching (DP: Dynamic Program)
The method of absorbing each of the ng) is conventionally used.

FIG. 5 shows an outline of a conventional word voice recognition device. In this device, the spoken voice is input from the voice input unit 1 and given to the acoustic analysis unit 2 as a voice signal. The acoustic analysis unit 2 calculates the characteristic parameter of the voice and obtains it as an input pattern.

On the other hand, in the word standard pattern storage unit 3, a plurality of word standard patterns obtained from learning sample data by a clustering method for each word registered in advance.
The word matching unit 4 calculates the degree of similarity by DP matching between the input pattern and the standard word pattern, and the determination unit 5 determines the word recognition result from the similarity. Was.

[0005]

However, in this type of apparatus, the number of word standard patterns increases in the case of increasing the recognition target vocabulary and dealing with unspecified speakers.

Therefore, in order to reduce the calculation amount and provide a word voice recognition method suitable for recognizing the voice of an unspecified speaker, the time series of the feature parameters of the input voice and the feature point detecting voice pattern are dynamically planned. A method of detecting the position of the feature point of the input voice by matching using the method, and matching the feature pattern of the input voice at the detected feature point with the standard feature pattern to convert the input voice into a symbol sequence of the feature pattern. Has been proposed (JP-A-59-77500). That is, in this apparatus, in order to reduce the number of DP matching at the time of matching, the DP matching is used only for the alignment of the feature points, the detailed matching is not performed by the DP matching, and the feature points obtained by the alignment are A detailed collation is performed by identification. Therefore, in order to deal with a large number of speakers, it is necessary to align the feature points with a plurality of standard patterns, and there is a problem that the processing amount of DP matching cannot be increased.

As described above, the conventional apparatus has a problem in that the calculation amount of DP matching in the word matching unit increases and the time required for the recognition processing is long.

The present invention has been made in view of the above circumstances, and an object of the present invention is to provide a word voice recognition device capable of identifying similar words of an unspecified speaker at high speed and with high accuracy.

[0009]

A word voice recognition apparatus of the present invention detects a spoken voice and inputs it as a voice signal, and a feature parameter based on a waveform of the input voice signal. And a characteristic point extracting means for extracting a characteristic point time series by detecting characteristic time points at which the norm of the amount of change in spectrum is maximum and minimum from the time series of characteristic parameters. And a word standard pattern storage means for storing a standard word pattern created in advance from learning sample data of word voice, and a similarity between the extracted feature point time series and the standard word pattern. And a determination means for outputting a recognition result from the word similarity. (Here, the norm means the sum of absolute values.) That is, in the present invention, a characteristic frame is extracted as a characteristic point without loss of phoneme information, and this is collated by so-called thinning processing. . In this extraction, the speech data is regarded as a trajectory in the feature space, and static information of the stationary part (position of the space ... Spectral feature amount) and dynamic information of the transient part (movement from space to space ... Spectral change feature quantity) and stationarity and transientness (continuity in a certain space ... Spectral change quantity: sum of absolute values of spectrum change feature quantity (norm)) are used as feature parameters. The maximum and minimum of the norm are detected, the point of time is set as a feature point, parameters of only the feature point are extracted, and the voice data is approximated by a time series of feature points composed of these feature parameters. Then, only these characteristic frames are collated by DP matching.

In the acoustic analysis means, it is desirable that the characteristic parameter used for time series conversion is the LPC cepstrum coefficient, and the time change amount (Δ) of the LPC cepstrum coefficient.
Cepstrum) and Δ cepstrum norm, that is, the dynamic scale. Then, the feature point extraction means sets the time point at which the maximum value and the minimum value of the norm are taken from the LPC cepstrum coefficient, the amount of time change of the LPC cepstrum coefficient (Δ cepstrum) and the norm of the Δ cepstrum as a feature point, and at this feature point , Feature parameters are extracted.

[0011]

In this device, when the uttered voice is detected and input, the input voice is analyzed and converted into the time series of the characteristic parameters, the characteristic time point is detected from the time series of the characteristic parameters. Then, the feature point time series is extracted, and the degree of similarity is calculated by comparing this with the word standard pattern, which reduces the processing amount and reduces the storage capacity required for the standard pattern. Reductions can be made. Further, in this device, continuous DP matching is possible without having to cut each word, so that it is easy to apply to continuous speech. According to the apparatus of the present invention, feature points can be properly extracted in this manner without loss of phonological information, and recognition can be performed at high speed and with high accuracy.

Further, preferably, as the characteristic parameter,
The LPC cepstrum coefficient, the time variation of the LPC cepstrum coefficient (Δ cepstrum), and the norm of the Δ cepstrum (dynamic scale) are used. Here, the maximum of the dynamic scale corresponds to the dynamic information of the transient part, and the minimum corresponds to the static information of the stationary part. Therefore, by extracting the maximum and the minimum of this dynamic scale as feature points, Maintain phonological information,
It is possible to extract only the most characteristic partial data, and it is possible to obtain a qualified output with a small amount of data processing.
It becomes possible to perform recognition with higher accuracy.

As the characteristic parameters, the LPC cepstrum coefficient, the amount of time change of the LPC cepstrum coefficient (Δ cepstrum), and the norm of the Δ cepstrum (dynamic scale) are used. , Three having the relationship of the sum of the absolute values of the change amounts may be used. For example, as the spectrum, the output from the band pass filter, the output from the FFT analysis, or the like may be used.

Then, DP matching is performed to compare the similarity between the extracted characteristic parameter series and the word standard pattern, the similarity is calculated, the most similar standard pattern is determined, and the recognition result is output. Therefore, high-speed and high-accuracy voice recognition can be performed.

[0015]

Embodiments of the present invention will now be described in detail with reference to the drawings.

FIG. 1 is a schematic block diagram showing an embodiment of the present invention. This device uses L as a characteristic parameter.
A feature point extraction unit 13 for extracting the maximum and minimum of the dynamic scale as feature points by using the time variation of the PC cepstrum coefficient, the LPC cepstrum coefficient (Δ cepstrum) and the norm of the Δ cepstrum (dynamic scale) is provided. It is characterized by having done. In addition, the word standard pattern storage unit 14 extracts all the maximum and minimum dynamic scales as feature points from the learning sample data of the word voice uttered by many speakers in advance by the same method. The word is created and stored as a standard pattern. Other parts are configured similarly to the conventional device shown in FIG.

That is, this apparatus includes a voice input section 11 composed of a microphone or the like so as to detect a voice and output it as a voice signal, an acoustic analysis section 12 for extracting a characteristic parameter from the voice signal, and an extraction. The maximum and minimum of the dynamic scale are obtained from the extracted feature parameters, and the feature point extracting unit 13 for extracting the maximum and the minimum of the dynamic scale, the word standard pattern storage unit 14, and the parameter time series at the time point extracted by the feature point extracting unit 13 are extracted. (Feature points) and word standard pattern storage unit 14
The word collating unit 15 that collates the stored data to find the similarity.
And a determination unit 16 that determines the input word voice from the word similarity obtained by the word matching unit 15 and outputs the recognition result.

Here, the word standard pattern storage unit 14 stores the standard pattern created by the above-mentioned feature point extraction method from the learning sample data of the spoken word voice in advance.

Next, a specific operation of the voice recognition device according to the embodiment of the present invention will be described.

The voice input unit 11 detects a spoken voice and supplies it as a voice signal to the acoustic analysis unit 12. Then, the acoustic analysis unit 12 extracts characteristic parameters from the given audio signal. This is to obtain the characteristic parameter time series of the input voice as shown in FIG. 2 by performing linear prediction analysis, cepstrum analysis, etc. on the given voice signal. FIG. 2 is an output example of the acoustic analysis unit 12 for a word voice spoken as "intermediate". The parameters to be extracted are the cepstrum coefficient (LPC cepstrum coefficient) by the linear prediction model,
Δ which is a regression coefficient for each dimension of the LPC cepstrum coefficient
It is a dynamic measure corresponding to the norm of the LPC cepstrum coefficient and the ΔLPC cepstrum coefficient. In FIG. 2, the LPC cepstrum coefficient is converted into spectral intensity and displayed. These characteristic parameter time series are output to the characteristic point extraction unit 13.

FIG. 3 is an explanatory diagram for explaining the processing of the feature point extraction unit 13. The characteristic parameter time series representatively displays a spectrogram converted from the LPC cepstrum coefficient. The feature point extraction unit 13 first pays attention to the dynamic scale among these feature parameters, and detects the time points at which the dynamic scale becomes maximum and minimum. In FIG. 3, the maximum point is the transitional part in the voice, which is the position of the consonant part such as / t / which is a plosive sound, and the inter-phoneme transition part such as / r / and / i /. In addition, the time point of the minimum is the stationary part in the voice, and is the position of the vowel part such as / i / or the consonant part such as the frictional sound such as / ts / which continues in time. Therefore, the minimum and maximum values of the dynamic scale are positions at which phoneme information is left at the phonemes or the transition parts between phonemes. Using this time point as a feature point, parameters of only the feature points are extracted from the feature point parameter time series to form the feature point time series. The characteristic parameters extracted as the characteristic points are the LPC cepstrum coefficient, the ΔLPC cepstrum coefficient, and the norm of the ΔLPC cepstrum coefficient. Then, this feature point time series is given to the word matching unit 15.

On the other hand, the word standard pattern storage unit 14 extracts the LPC cepstrum coefficient and the ΔLPC cepstrum coefficient for all words from the learning sample data of the word speech spoken by a large number of speakers in advance by the above-described feature point extraction method. And a dynamic measure of the ΔLPC cepstrum coefficient,
It is obtained by extracting the maximum and the minimum of this dynamic measure and storing them as a standard pattern.

Then, the word matching unit 15 performs DP matching between the feature point time series pattern and the word standard pattern, and outputs the word similarity. Figures 4 (a) and 4 (b) show a comparison of the amount of calculation required for this DP matching. 4 (a) and 4 (b) respectively show the DP calculation amount in the conventional method and the DP calculation amount in the method of the present invention, in which the DP calculation amount corresponds to the area within the matching window. From these comparisons, it can be seen that the present invention significantly reduces the calculation amount.

The word similarity obtained by the word collating section 15 is calculated in this way, and is output to the judging section 16.

Then, the judgment unit 16 selects the word having the maximum word similarity from all the words, and judges the result from the threshold value. In this way, high-speed and high-accuracy voice recognition can be performed.

The present invention can be appropriately modified without departing from the spirit of the invention, and can be applied to various fields such as online handwritten character recognition.

[0027]

As described above, according to the present invention, the speech recognition of a large vocabulary or an unspecified speaker is performed by handling the parameter time series of only the characteristic points of the uttered speech by DP matching. In, the recognition speed can be improved without deteriorating the recognition ability.

[Brief description of drawings]

FIG. 1 is a schematic block diagram of an embodiment of the present invention.

FIG. 2 is a diagram showing an output example of an acoustic analysis unit 12 in one embodiment of the present invention.

FIG. 3 is a diagram showing an output example of a feature point extraction unit 13 in one embodiment of the present invention.

FIG. 4 is a comparison diagram of DP calculation amounts between the conventional method and the present invention.

FIG. 5 is a schematic block diagram showing a conventional word voice recognition device.

[Explanation of symbols]

 1 Speech Input Section 2 Acoustic Analysis Section 3 Word Standard Pattern Storage Section 4 Word Matching Section 5 Judgment Section 11 Speech Input Section 12 Acoustic Analysis Section 13 Feature Point Extraction Section 14 Word Standard Pattern Storage Section 15 Word Matching Section 16 Judgment Section

Claims (2)

[Claims] 1. A voice input means for detecting a spoken voice and inputting it as a voice signal pattern, and an acoustic analysis means for converting into a time series of characteristic parameters based on the waveform of the input voice signal pattern. From the time series of feature parameters, the feature point extracting means for detecting the time points at which the norm of the amount of change in the spectrum becomes maximum and minimum as feature points, and extracting the feature point time series, and learning sample data of the word speech in advance. The word standard pattern storage means for storing the standard word pattern created from the data, the word matching means for calculating the similarity between the extracted feature point time series and the standard word pattern, and the word matching means A word voice recognition device, comprising: a determination unit that outputs a recognition result from the obtained word similarity. 2. The acoustic analysis means is configured to use an LPC cepstrum coefficient, an amount of time change of the LPC cepstrum coefficient (Δ cepstrum) and a norm of the Δ cepstrum as characteristic parameters. 1. The word voice recognition device described in 1.