JPS63191199A - Voiced plosive consonant identifier - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed Description of the Invention] [Summary] The present invention provides a device for identifying voiced plosive consonants that targets unspecified speakers. When performing adaptation, speaker analysis can be performed stably by using the spectral time series obtained by performing voice analysis in the reverse temporal direction from the subsequent vowel side, rather than using the rupture point as the base point, where detection accuracy is unstable during learning. It is designed so that it can be changed.

[Industrial application field]

The present invention relates to a method for mutually identifying voiced plosive consonants in a speech recognition device. Speech Ga devices, especially monosyllable recognition devices, can basically form all Japanese words with 68 monosyllables, so they can be used to input sentences instead of keyboard input, and are currently being put into practical use as voice word processors. is being measured. However, in monosyllabic speech recognition, differences in the characteristics of the consonant part at the beginning of ff# account for a large proportion of the difference, and the methods conventionally used in word speech recognition devices cannot capture these characteristics. Especially speaking! !
It is difficult to distinguish between consonant groups that have very similar structures, and a high-precision discrimination method is needed. On the other hand, speech recognition devices for non-specific speakers have been put into practical use for a small number of words, but it is not easy to absorb the variable factors of speakers.

It is very difficult to obtain a high identification rate. To deal with this, a method has been adopted in which a small number of learning utterances are requested from the user in advance, and adaptation to the speaker is performed based on the data.

[Conventional technology]

Conventional methods for identifying voiced plosive consonants generally include a method that uses a plosive part spectrum time series immediately after the plosive point, or a method that uses a spectral time series from the plosive point to the following vowel. In addition, in the voiced plosive consonant identification method for unspecified speakers, a standard pattern is created in advance from voice data uttered by multiple speakers and compared against it in a dictionary.

A method of judgment is used.

By the way, in a voiced plosive consonant identification method targeted at unspecified speakers, in order to increase the identification rate, the user may be requested to utter a small number of utterances in advance to perform nine-speaker adaptation. Conventionally, a spectrum or spectral time series, which is learning data, was obtained from input audio based on the point of rupture. FIG. 3 shows a conventional example in which a spectral time series is used.

In the figure, 11 is a microphone, 12 is an A/D converter, and 13
is an A/D data memory, 14 is a rupture point detection section, 15 is a frequency analysis section, 16 is a spectral time series memory for learning,
17 is a speaker adaptation unit, and 18 is a dictionary for unspecified speakers.

The learning audio input from the microphone 11 is as follows.

Analog-to-digital conversion is performed by the A/D converter 12 and stored in the A/D data memory 13.

From the audio signal stored in the memo+713, the time point at which the plosive phenomenon peculiar to plosive sounds occurs (the rupture point) t - the rupture point detection unit 1
4 to detect. A frequency analysis unit 13 performs spectrum analysis on the subsequent vowel side based on the point of rupture, and obtains a spectrum time series. This spectral time series is stored in the learning spectral time series memory 16. Speaker adaptation unit 1
In step 7, speaker learning of the standard butter/ stored in the dictionary 18 for unspecified speakers is performed based on this spectrum time series.

[Problem that the invention seeks to solve]

Conventionally, voiced plosive consonant identification methods (
In the speaker adaptation method in which the spectral sequence from the plosive point to the following vowel is used as a feature quantity,
Based on a small number of sounds uttered by the user in advance, the point at which the tzu ruptures is detected, and the spectral time series that can be applied to the subsequent vowel from that point to the point at which the tzu ruptures is used for learning and speech using a dictionary for non-specific speakers. It is adapted to people. However, the automatic detection rate at the point of rupture varies greatly depending on the speaker and utterance. Therefore, extracting the spectral time series used during learning based on the burst point will significantly impair the reliability of the training data, and it will not be possible to improve the classification rate through learning, that is, to accurately adapt to nine speakers. .

[Means for solving problems]

@Figure 1 is a diagram showing the principle of the voiced plosive consonant identification device according to the present invention.
2 is an analysis position setting circuit, 3 is a frequency analysis circuit, 4 is a dictionary for unspecified speakers, 5 is a burst point detection circuit, and 6 is a speaker adaptation circuit.

In the present invention, first, the subsequent vowel rise time detection circuit 1 detects the subsequent vowel rise time with respect to the audio signal input from the microphone and A/D converted. Then, based on that point in time, the zero frequency analysis circuit 3 performs frequency analysis in the reverse time direction according to the analysis position set by the analysis position setting circuit 2. The rupture point detection circuit 5 detects a time point having a spectrum near the rupture point by comparing it with a standard pattern of buzz buzz (or silence) stored in advance in the dictionary 4 for unspecified speakers. Thereafter, the speaker adaptation circuit 6 performs speaker adaptation of the speaker-independent dictionary using the obtained spectrum from the rising point of the subsequent vowel to the rupture point.

[Effect]

In the present invention, the spectral time series obtained from the learning speech is obtained by performing spectrum analysis in the reverse time direction with reference to the rising point of the subsequent vowel, and the detection accuracy is unstable for each of the nine speakers and utterances. It is possible to obtain more stable learning data than the conventional spectral time series, which is obtained based on the point of rupture.

Speaker learning is performed accurately.

〔Example〕

FIG. 2 is a schematic diagram of a voiced plosive consonant identification device for all unspecified speakers according to an embodiment of the present invention.

In the figure, a learning discrete monosyllabic voice (voiced plosive) inputted from a microphone 21 is converted from analog to digital by an A/D converter 22 and stored in an A/D data memory 23. A subsequent vowel rise time detection circuit 24 detects the rise time of the subsequent vowel from the audio signal stored in the A/D gator memory 23. That is, the audio power calculation circuit 25 calculates the audio power temporal series. Now, if the audio signal is y (t), the audio power P is P=Σy''(t
) (N is the frame length).

According to the above definition, an audio power temporal sequence is determined with a frame period M. The audio power temporal series is stored in the audio power temporal series memory 28. Next.

Based on the speech power temporal sequence, the steady part of the following vowel is detected by the vowel steady part detection circuit 26. The processing here is
In the speech power temporal series, a point in time when the power is large and the variation thereof is less than or equal to a certain color value is detected, and is determined as the starting point in the steady part of the vowel.

Next, the vowel rising time detection circuit 27 detects the vowel rising time. Here, the time point at which the power decreases by a certain predetermined power value from the voice power value at the start of the steady part of the vowel is determined, and that time point is defined as the vowel's rising point. Next, frequency analysis is performed in the FFT calculation circuit 30 in the reverse time direction based on the rising point of the subsequent vowel detected by the 9 circuit 24 and according to instructions from the analysis position setting circuit 29. The inter-bang distance between the resulting spectrum and the buzzer and silent standard spectra stored in the speaker-independent dictionary 34 is calculated by the inter-bang distance calculation circuit 32. The buzz bars and silent standard spectrum bangs stored in the speaker-independent dictionary 34 are created in advance from the voices of many speakers. Based on this inter-bang distance, the rupture part determination circuit 33 determines whether the spectrum is the rupture part, stores it in the learning spectrum time series memory 35, and if the rupture part has not been reached. returns to the analysis position setting circuit 29 again,
Continue frequency analysis in the reverse time direction. If the rupture part is reached, proceed to the ninth step. In the next step, first, the data selection circuit 36 selects the spectrum at the vowel rise time, the spectrum near the center of the transition part, and the spectrum time series at the rupture time from the spectrum time series stored in the learning spectrum time series memory 35. . The spectral distance calculation circuit 37 calculates the spectral distance between this spectral time series and the standard spectral time series bang stored in the speaker-independent dictionary 34. Note that the utterance content of the learning utterance is already known, and is instructed by the speaker through the input unit 39. Further, the standard spectrum time series patterns for each category stored in the speaker-independent dictionary 34 are created in advance from a large number of speakers (for example, representative patterns selected by a clustering method, etc.). Also, bl in the dictionary for unspecified speakers 34.

b2. b3 are the spectrum at the rupture point, the spectrum at the center of the transition part, and the spectrum at the vowel rise point, respectively. As a result of the spectral distance calculation by the spectral distance calculation circuit 37, the flag area (Fl) in the dictionary 14 is detected for the standard spectral time series button whose spectral distance is less than or equal to the set threshold value.

-・) Raise a flag. The standard spectrum time series with this flag set will be a dictionary adapted to the speaker who is using it. At the time of recognition, the dictionary selection circuit 38 refers to the flag and selects a dictionary item adapted to the speaker used for identification.

〔Effect of the invention〕

According to the present invention, instead of using the unstable rupture point for each category of two-voiced sounds as a reference, it is possible to obtain a learning spectral time series in the reverse time direction using the relatively stable rising point of the subsequent vowel as a reference. It is possible to create stable learning data, and has a great effect on speaker adaptation for the speaker using it.

[Brief explanation of the drawing]

Figure 1 is a diagram showing the principle of the present invention. FIG. 2 is a block diagram of a voiced plosive consonant identification device according to an embodiment of the present invention. FIG. 3 shows an example of the configuration of the prior art. In FIG. 1, 1 is a subsequent vowel rising point detection circuit;
2 is an analysis position setting circuit, 3 is a frequency analysis circuit, 4 is a dictionary for unspecified speakers, 5 is a burst point detection circuit, and 6 is a speaker adaptation circuit.・'Minki Kinmei Q Hara Tanuki Pro Riff Diagram % 1 Figure 2 Tosai Cho n Composition Example Figure 3

Claims (1)

[Scope of Claims] Subsequent vowel rise detection means (1) for detecting the rise point of a subsequent vowel regarding an inputted voiced plosive, and analysis position setting that sets an analysis position based on the rise point of the subsequent vowel. means (2); frequency analysis means (3) for performing frequency analysis in the reverse time direction according to the analysis position set above; and storing spectrum time series information including silence and bus bar spectra extracted from multiple speakers in advance. Voiced information inputted by comparing the information output from the speaker-independent dictionary (4) and the frequency analysis means (3) with the silence and busbar information in the speaker-independent dictionary (4). A plosive point detection means (5) for detecting the plosive point of a plosive, and a speaker who performs speaker adaptation processing of the speaker-independent dictionary using the obtained spectrum from the rising point of the subsequent vowel to the plosive point. A voiced plosive consonant identification device comprising: an adaptation means (6).