JPH02226300A - Phoneme section information generating device - Google Patents

Abstract

PURPOSE:To improve segmentation accuracy and to accurately detect the transition section of a vowel from a low-frequency voice which has small power variation by using a formant as a parameter for a phoneme section, finding variation in its time direction and extracting the border between a stationary part and a transient part as a feature point. CONSTITUTION:The voice signal from a microphone 1 is inputted to an A/D converting circuit 4 through an amplifier 2 and a low-pass filter 3 and converted into, for example, an 12-bit digital signal of 12.5kHz in sampling frequency, which is supplied to an acoustic analyzing means 5. The means 5 consists of a transient part detection parameter generating means 51 which has a band-pass filter bank, a logarithmic power detecting means 52, a zero-cross rate arithmetic means 53, a PARCOR coefficient arithmetic means 54, a power spectrum gradient arithmetic means 55, a format detecting means 56 and a voice basic cycle detecting means 57. Thus, the parameter obtained by the means 5 is supplied to a phoneme recognizing means 8 and respective parameters of the means 51 - 56 are supplied to the feature point detecting means 61 of a 1st segmentation means 6.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a phoneme segment information forming device, and more particularly to a phoneme segment information forming device for forming phoneme segment information for dividing a phoneme segment of input speech.

[Summary of the invention]

In a phoneme interval information forming device, the present invention includes an acoustic analysis means that performs acoustic analysis of input speech and obtains changes in formants in the time direction as parameters for phoneme intervals; By comprising a means for extracting the boundary with the transitional part as a feature point, and a means for obtaining phoneme interval information using the information on the feature points, it is possible to further improve the accuracy of segmentation. be.

[Conventional technology]

In the case of continuous speech and large word speech recognition, phonological recognition is the basis. In this phoneme recognition, it is necessary to perform so-called segmentation, which divides the input speech to be recognized into phoneme sections.

For example, when pronouncing the word [SU(SU)J], the speech waveform can be phonetically divided into a consonant "S" and a vowel "U".

As a method for this segmentation, there has conventionally been a method of comparing the voice power or zero-crossing rate with a threshold value to find division points (section boundaries).

However, it has been difficult to accurately segment phoneme intervals using such a method due to the difficulty in setting a threshold value.

Therefore, the applicant of the present application has proposed a technique for accurately performing segmentation of phonetic intervals in Japanese Patent Application No. 62-323307.

[Problem to be solved by the invention]

By the way, formant is generally known to be a useful parameter in speech recognition.

However, since the technique of the above-mentioned prior application does not use formants as a phoneme interval parameter for performing segmentation, there is a problem that segmentation may not be performed with high accuracy, and an improvement has been desired.

Therefore, an object of the present invention is to provide a phoneme segment information forming device that can perform more accurate segmentation by adding formants to phoneme segment parameters.

[Means for Solving the Problems] The present invention provides an acoustic analysis means that performs acoustic analysis of input speech and obtains changes in the formant in the time direction as a parameter for the sound interval, and a stationary The device is configured to include means for extracting the boundary between the transition region and the transition region as a feature point, and means for obtaining phoneme interval information using the information on the feature point.

[Effect]

Obtain temporal changes in formants by acoustic analysis of input speech.

Next, regarding the change in the formant in the time direction, the boundary between the stationary part and the transient part is extracted as a feature point. When a formant change point is detected at approximately the same position as a phoneme boundary candidate determined from conventional parameters, the conventional phoneme boundary candidate is used as is. If a formant change point is detected at a position that is not detected using conventional parameters, the formant change point is adopted as a new phoneme boundary candidate and a phoneme boundary feature is attached to it.

In this way, by using the formant as a phoneme interval parameter, it is possible to perform accurate segmentation.

〔Example〕

An embodiment of the present invention will be described below with reference to FIGS. 1 to 5.

FIG. 1 shows an example of a speech recognition device according to the present invention.

The audio signal from microphone 1 is sent to amplifier 2 and low
The signal is supplied to an A/D conversion circuit 4 via a pass filter 3. The above-mentioned audio signal is converted into a 12-bit digital audio signal by the A/D conversion circuit 4 at a sampling frequency of, for example, 12.5 K11z.

This digital audio signal is supplied to acoustic analysis means 5.

The acoustic analysis means 5 includes a transient detection parameter generation means 51 having a bandpass filter bank, a logarithmic power detection means 52 for detecting audio power, a zero cross rate calculation means 53, and a primary a calculating means 54 for calculating the Percoll coefficient, a calculating means 55 for calculating the slope of the power spectrum, a formant detecting means 56 having a band-pass filter bank (not shown) for determining changes in the formant in the time direction, and detecting the fundamental period of the voice. Means 57 is provided.

First, generation of transient detection parameters will be explained.

The transient detection parameter is used to detect the transient nature and stationarity of the input audio, and this transient detection parameter is used to calculate the amount of change in the audio spectrum by the sum of the variance within the block in the time direction of each channel (frequency). is defined as That is,
The gain of the audio spectrum 5t(n) is normalized using the average value S avg (n) shown below in the frequency direction.

Here, i indicates a channel number, and q indicates the number of channels (number of bandpass filters). Furthermore, although the information of each channel of the q channel is sampled in the time direction, a block of information of the q channel at the same time is called a frame, and n indicates the number of the frame used for recognition.

Gain normalized speech spectrum S i (n)
is S i (n) −S i (n) −S avg (
n)------(2).

The transient detection parameter T (n) is the sum of H frames before and after that frame (2M+1) [n-gate, n+
[river] is defined as the sum of the temporal variance of each channel within the block.

Here, is the average value in the time direction within the block of each channel.

In practice, changes near the center of the (n-M+n+M) block tend to pick up sound fluctuations or noise, so they can be removed from the calculation of the transient detection parameter T (n), and equation (3) is as follows: transformed into.

T (n) = 2q(M-m+1) And in equation (5), a=1, H=28, m=
3. The transient detection parameter T(n) is determined as q-32. For example, in the case of the input voice [morning (asa) J], the transient detection parameter T (n
) is obtained.

Other parameters, such as the logarithmic power shown in Figure 2B, the zero crossing rate shown in Figure 2C, the first-order Percoll coefficient shown in Figure 2, and the slope of the power spectrum shown in Figure 2 Similarly to the transient detection parameter T(n), calculation of parameters such as detection and detection of the slope of the fundamental period shown in Fig. It is obtained by considering a line having a line -111 (5), sequentially moving this window one sample point at a time in the time direction, and performing calculations within each window. In addition, Fig. 2 F and J show the speech waveform of the input voice [asa J], Fig. 2 1 shows the formant transition, and Fig. 2 G and K show the waveform of the input voice [asa J], and Fig. 2 G and K show the waveform of the input voice [asa J]. An example of a phonological boundary candidate is shown below. Second
In the figure, F and J, G and K have the same content and are shown redundantly for convenience of comparison with other parameters.

Next, detection of a change in formant as a parameter in the time direction will be described.

Formants F1, F2, and F3 are extracted from the frequency spectrum of the vowel /a/ shown in FIG.
As shown in the figure, changes in the formant in the time direction are detected and output as parameters. Furthermore, Figure 4 shows the formant transition when the vowel /a/ changes to the vowel /i/, and Figure 5 shows the change from the vowel /a/ to the vowel /a/ via the consonant /m/. The formant transitions are shown.

Each parameter obtained from the acoustic analysis means 5 is supplied to the phoneme recognition means 8 as recognition processing parameters. Further, each parameter outputted from the means 51 to 56 is supplied to the feature point extraction means 61 of the first segmentation means 6 as a parameter for segmentation.

In the first segmentation means 6, -
Extract the characteristic points of boat fishing. In this example, the following seven types of feature points are used.

■ Rising point - the point where the flat part changes to an increasing direction ■ Falling point - the point where it changes to a decreasing direction and then becomes flat ■ Increased change point - the point where the rate of increase changes ■ Decrease change point - the decreasing rate ■ Peak point - the position of the peak ■ Positive zero crossing point - the point that intersects the zero level in the increasing direction ■ Negative zero crossing point - the point that intersects the zero level in the decreasing direction The feature point extraction means 61 extracts the feature points. The feature points are extracted for each parameter by referring to the feature point information from the information storage means 62.

An example of extracting a change point of formant as a parameter for a filN interval as a feature point will be described with reference to FIGS. 4 and 5. Note that the formant transition is expressed as follows.

(1) When a point of change in formant is detected in a vowel interval, vowel-vowel, t (point where formant change starts from the stationary part of the vowel, hereinafter abbreviated as l-V, T) v
owel, t-vowel (the point where the formant change ends and the stationary region begins, hereinafter abbreviated as V, T-V) (2
) When a point of change in formant is detected in a consonant interval, consonant-consonant, t (The point at which formant change starts from the stationary part of a consonant, hereinafter referred to as C-
(abbreviated as C0T) consonant, t-consonant (The point where the formant change ends and the constant part of the consonant begins, hereinafter abbreviated as C, TC) In Figure 4, between time tO and t1, formant F
There is almost no change in the frequencies 1 to F3, and the frequencies are determined to be in a steady state.

After time t1, the frequency of Holman F2 increases and the frequency of Holman 1-Fl decreases.

Therefore, the frequencies of Holman 1-F2 and Fl at time t1 are determined to be change points PFIO and PP20. This change point PFIO1PF20 is V-V based on the above notation.
.

It is written as T. This state continues between time points t1 and t2. Therefore, the period from time t1 to t2 is determined to be a transition period.

After time t2, the frequencies of formants F2 and Fl become stable. Therefore, formant F2 at time t2
, Fl is determined to be the changing point PFII, PP21. These change points PFII and PP21 are written as V and TV based on the above-mentioned notation. This state continues after time t2. Therefore, the period after time t2 is determined to be a steady portion.

Next, in FIG. 5, there is almost no change in the frequencies of the formants, F1, and F2 between time point 10 and time point tlO, and it is determined that this is a stationary region.

After time t1, the frequency of Holman) F2 increases and the frequency of Holman) Fl decreases.

Therefore, the frequencies of formants F2 and Fl at time t1' are determined to be the change point PFIO1PF20. This state continues between time points t1 and t2. Therefore, time tl~
The period t2 is determined to be a transition period.

After time t2, it enters a consonant period, and there is a tendency for power to concentrate in the low frequency range corresponding to formant F1, indicating that this is the stationary part of the consonant /m/. Therefore, the frequencies of formants F2 and Fl at time t2 are determined to be change points PFII and PP21. This change point pp.
Based on the above notation, i1 and PP21 are C, T-C
It is written as This state continues from time t2 to time t3. Therefore, the period from time t2 to L3 is determined to be the constant part of the consonant.

After time t3, the frequency of Holman) F2 decreases and the frequency of Holman) Fl increases.

Therefore, the frequencies of Holman's F2 and Fl at time t3 are determined to be change points PF12 and PP22. This state continues between time points t3 and t4. Therefore, time t3 to t
The period between 4 and 4 is determined to be a transition section. This change point PF12, P
P22 is written as C-COT based on the above-mentioned notation.

After time t4, the frequencies of Holman) F2 and Fl become stable. Therefore, formant F2 at time t4
, Fl are determined to be changing points PF13 and PP23. This state continues after time t4.

Therefore, the period after time t4 is determined to be a steady portion.

FIG. 2 shows the formant F detected in this way.
Change points from 1 to F4 are shown.

In this way, formant is used as a parameter for phoneme interval,
Since feature points can be extracted, more accurate segmentation is possible.

Among the parameters shown in FIGS. 2A-B and I, the positions indicated by vertical lines in the time axis direction are the positions of each feature point.

Each parameter obtained from the first segmentation means 6 and to which feature points are attached is transmitted to the second segmentation means 7
is supplied to

The second segmentation means 7 includes a feature point integration processing means 71, a phoneme boundary feature detection means 72, a feature point integration information storage means 73, and a phoneme boundary feature information storage means 74.

Since the feature points obtained by the first segmentation means 6 have positional deviations, undetected points, etc. for each parameter, the feature point integration processing means 71 refers to the feature point integration information from the feature point integration information storage means 73. The feature points of each parameter are summarized to determine phoneme boundary candidates. Note that the feature point integration information is information about which parameter's feature point should be prioritized.

When a change point PF obtained by formant detection is detected at approximately the same position as a phoneme boundary candidate obtained using other parameters, the feature point integration processing means 71 uses the phoneme boundary candidate obtained using other parameters. adopt. On the other hand, when the change point PF obtained by formant detection is a phonological boundary candidate that has not been detected from other parameters, the change point PF obtained from the formant is adopted as a new phonological boundary candidate, and segmentation I do.

The phoneme boundary feature detection means 74 determines the phoneme boundary feature of each phoneme boundary candidate. In this example, the following 12 types of phoneme boundary features are used.

■ Rise from silence (S-R) ■ Consonantity → Vowelness (C-V) ■ Vowelness → Vowelness (V-V) ■ Vowelness → Vowel transition (V-V, T) ■ Vowelness Transitional part → Consonantity (V, T-C) ■ Consonantity → Vowel transitional part (C-
V, T) ■Transitional part of vowel → vowel character (V, T-V) ■Falling to silence (F-3) [Phase] Sound → Silence (S-S) [Phase] Consonant character → Consonant character (C-C) ■Consonantity → Consonant transition part (C-C, T) @ Consonant transition part → Consonantity (C, T-C) Phonological boundary feature information storage means 74
These 12 types of phoneme boundary feature information are stored in , and the phoneme boundary feature detection means 72 detects the phoneme boundary feature of each phoneme boundary candidate by referring to the information from the phoneme boundary feature information storage means 74. . As a result, as shown in FIGS. 2G and 2K, the above-mentioned nine types of phoneme boundary features (■ to ■) are shown in the order of ■ to ■ in the vicinity of the vertical line of the phoneme boundary candidate.

The second segmentation means 7 obtains phoneme boundary candidate information and its phoneme boundary feature information as phoneme interval information. This phoneme segment information is then supplied to the phoneme recognition means 8.

The phoneme recognition means 8 uses the parameters from the acoustic analysis means 5 as parameters for recognition processing, and executes phoneme recognition while referring to the phoneme interval information from the second segmentation means 7. As a result, phoneme recognition for each phoneme section is performed, for example, as shown in FIGS. 2F and 2J, and the results are displayed.

The notation of this phonetic interval has the following meaning.

(1) - (a) Constant part of vowel /a/ (2) - (as) Transition part of vowel /a/ to consonant /s/ (
3) - (s) Constant part of / to fricative consonant (4) - (sa) Transition part of vowel /a/ from consonant /s/ (5) - (a) Stationary part of vowel /a/ and phoneme Recognized phonetic symbols are obtained from the recognition means 8, and are supplied to continuous speech and large word speech recognition means at the subsequent stage.

In this way, formant is used as a parameter for phoneme interval,
Since feature points are extracted, phonological boundary candidate information and phonological boundary feature information can be extracted from formant transitions, and it is possible to accurately detect phonological intervals of semi-vowels where the formant changes slowly. It is possible to accurately detect vowel-to-vowel transition sections.

In this embodiment, an example is explained in which the hardware is used, but the first and second segmentation means 6.
7. The calculation part of the acoustic analysis means 5, the phoneme recognition means 8, etc. may be realized by a computer.

〔Effect of the invention〕

According to this invention, the formant is used as a parameter for phoneme interval, the change in the formant in the time direction is obtained, and the boundary between the stationary part and the transient part is extracted as a feature point, so the segmentation accuracy is higher than [-1]. There is an effect that it can be done. Furthermore, it is possible to accurately detect phonetic intervals of semi-vowels (such as ya, yu, yo, kya, etc.) whose formants are slowly changing, thereby improving the phoneme recognition rate. Furthermore, it is possible to accurately detect a vowel-to-vowel transition section with little power change, and it is possible to improve the recognition rate of vowel chains.

[Brief explanation of the drawing]

Fig. 1 is a block diagram showing an embodiment of the present invention, Fig. 2 is a waveform diagram for explaining each embodiment, Fig. 3 is an explanatory diagram showing formants, and Figs. 4 and 5 are formant transitions, respectively. FIG. Explanation of main symbols in the drawings 5: acoustic analysis means, 56: formant detection means, 6: first segmentation means, 7: second segmentation means. Procedural amendment written by Ku-yen-b (Ka-style, June 1989, 1) Commissioner of the Patent Office Yoshi 1) Takeshi Moon 1, Indication of the case 1999 Patent Application No. 46969 2, Name of the invention Phonology Section information forming device 3, relationship with the case of the person making the amendment Patent applicant address 6-7-35, Kitashina, Tokyo Parts Store Name (21
8) Sony Corporation Representative Director Norio Ohga 4, Agent 170 Address 1-48-10 Higashiikebukuro, Toshima-ku, Tokyo May 30, 1999 (Delivery) 6. Drawings of the specification subject to amendment From column 7 for a brief explanation of the amendment, from line 11 to line 12 on page 19 of the statement of contents of the amendment,
The phrase "The third figure is a formant" should be corrected to read "The third figure is a formant."

Claims (1)

[Scope of Claims] Acoustic analysis means that performs acoustic analysis of input speech and obtains changes in formants in the time direction as parameters for phoneme intervals;
A phoneme interval information forming device comprising means for extracting a boundary with a transitional part as a feature point, and means for obtaining phoneme interval information using information on the feature point.