JPH01165000A - Vocal sound section information forming apparatus - Google Patents

Abstract

PURPOSE: To correctly segment a phoneme section by providing an acoustic analyzing means, a feature point extracting means, and a phoneme section information output means. CONSTITUTION: The acoustic analyzing means 5 takes an acoustic analysis of an input speech and finds parameters for phoneme sections and the feature point extracting means 61 extracts feature points such as a rise, a fall, and a zero crossing as to the parameters for the phoneme sections. The phoneme section information output means 7 finds phoneme section border candidates and phoneme border features (e.g. from consonant to vowel, from vowel to vowel, and rise from voiceless sound) of the respective border candidates from information on the feature points of the respective parameters and outputs those phoneme section border candidates and phoneme border candidates as vocal sound section information. Consequently, the accurate phoneme section information can easily be obtained.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to an apparatus for forming phoneme segment information, which is important for dividing phoneme segments of input speech in speech recognition, particularly sound B recognition.

[Summary of the invention]

This invention uses information on feature points such as rising, falling, and peak points of each of multiple types of phonetic segment parameters obtained through acoustic analysis of input speech to identify phonetic segment boundary candidates, rising from silence, consonant-vowel, The phonetic boundary features of each boundary candidate such as vowel-vowel are determined and these are used as phonetic interval information, making it possible to accurately and efficiently determine the phonetic interval.

[Conventional technology]

In the case of continuous speech and large volume speech recognition, phonological recognition is the basis. For this phoneme recognition, it is necessary to divide the target input speech into phoneme sections.

For example, when pronouncing the word "su", the speech waveform can be phonetically divided into a consonant "S" and a vowel "U".

Divide into this phonological interval (hereinafter referred to as segmentation)
Conventionally, the method used is to compare the power of the voice, the zero-crossing rate, etc. with a threshold value and find the dividing point (interval boundary).

[Problem that the invention seeks to solve]

However, when phonological segmentation is performed simply by comparing speech power or zero crossing rate with a threshold value, it is difficult to accurately segment the phonological interval, coupled with the difficulty of setting the threshold value. It was difficult.

The present invention aims to propose a device that can provide information that enables more accurate segmentation of phoneme intervals.

[Means for solving problems]

In this invention, acoustic analysis means, feature point extraction means, and phoneme interval information output means are provided.

[Effect]

The acoustic analysis means acoustically analyzes the input speech and obtains parameters for a plurality of phoneme intervals.

The feature point extracting means extracts feature points such as rising edges, falling edges, and zero crossings for the phoneme interval parameters.

The phoneme interval information output means calculates phoneme interval boundary candidates and phoneme boundary features of each boundary candidate (for example, consonant → vowel, vowel → vowel, rise from silence, etc.) from the information on the feature points of each parameter, and The boundary candidates and phoneme boundary features are output as phoneme interval information.

〔Example〕

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

That is, an audio signal from a microphone (1) is supplied to an A/D converter (4) via an amplifier (2) and a band-limiting low-pass filter (3).
It is converted into a 12-bit digital audio signal at a sampling frequency of 2.5kHz. This digital audio signal is supplied to acoustic analysis means (5).

This acoustic analysis means (5) is, in this example, a transient detection parameter generation means (51) having a bandpass filter bank.
and logarithmic power detection means (52) for detecting voice power.
, zero cross rate calculation means (53), first-order Percoll coefficient calculation means (54) for checking the correlation between adjacent samples, and power spectrum slope calculation means l1
(55), and means (56) for detecting the fundamental period of voice.

The transient detection parameter is used to detect the transient nature and stationarity of the input audio, and this transient detection parameter defines the amount of change in the audio spectrum as the sum of the variance within the block in the time direction of each channel (frequency). be done.

That is, first, the gain of the audio spectrum S i (n) is normalized using the average value 3 av g (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 flight used for recognition.

Gain normalized speech spectrum S i (n)
is 5i(n)-5i(n)-Savg(n)
+・−r2).

The transient detection parameter T (n) is C n-M, which is the sum of M frames before and after the frame (2M+1.);
n+M] is defined as the sum of time-direction variances of each channel within the block.

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

In addition, in practice: nM, n2M] Changes near the center of the block are likely to pick up sound fluctuations and noise, so we decided to remove them from the calculation of the transient detection parameter T (n), and use Equation (3) is changed as follows.

In equation (5), a=l, M=28.
m=3. The parameter T (n) is found with q=32, and for example, in the case of the input voice "Asa J", it will be as shown in Figure 2 A. Note that Figure 2 G is the input voice waveform at this time. .

Other parameters are logarithmic power, zero cross L/-1
Similarly to the calculation of transient detection parameters, calculations for detecting the first-order Percoll coefficient, power spectrum slope, and fundamental period slope are based on a time width of M frames before and after a certain frame. Consider a window, move this window one sample point at a time in the time direction, perform calculations within each window,
Generate each parameter.

FIG. 2B shows the logarithmic power thus obtained, FIG. 2C shows the zero cross rate, the same first-order Percoll coefficient, and FIG. 2E shows the slope of the power spectrum.

Further, F in the figure shows the voice pitch, that is, the fundamental period of the voice.

The parameters obtained as described above are supplied from the acoustic analysis means (5) to the acoustic analysis means (8) as parameters for recognition processing. Further, each parameter from the means (51) to (55) is supplied as a segmentation parameter to the feature point extraction means (61) of the first segmentation means (6).

This first segmentation means (6) extracts characteristic points in order to obtain phoneme boundary candidates from the segmentation parameters. In this example, the following seven types of feature points are used.

(1) Rising point -1'- (2) Falling point -)-1 Feature point extraction method 1+ (61) extracts feature points for each parameter by referring to the feature point information from these feature point information storage means. In each parameter of FIGS. 2A to 2E, the positions indicated by vertical lines in the time axis direction are the respective feature point positions.

Each extracted feature parameter obtained from the first segmentation means (6) is supplied to the second segmentation means [7]. This 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 semi-immersed (74). ).

The feature points obtained by the first segmentation means (6) are:
Since each parameter has positional deviations, non-detections, etc., the feature point integration processing means (71) refers to the feature point information from the storage means (73), collects the feature points of each parameter, and selects phoneme boundary candidates. decide. The feature point integration information is information about which parameter's feature points should be taken with priority. An example of this phoneme boundary candidate is shown below the speech waveform in FIG. 2G by a vertical line in the time axis direction.

Further, the phoneme boundary feature detection means (72) determines the phoneme boundary feature of each phoneme boundary candidate. In this example, the following eight types of phoneme boundary features are used.

■Rising from silence (S-R) ■Consonantity → Vowelness (C-V) ■Consonance → Consonance (C-C) ■Vowelance → Vowelness (V-V) ■Falling to vowelness (V-F) ■Vowelality → Consonance <v-c> ■Falling to silence (F-3) ■Speech → Silence (S-3) These 8 are stored in the phoneme boundary feature information storage means (74). The phoneme boundary feature information of each type is stored, and the phoneme boundary feature detecting means (72) refers to the information from the storage means (74) to detect the phoneme boundary feature of each phoneme boundary candidate. Second
In the lower part of Figure G, near the vertical line of the phoneme boundary candidate, this phoneme boundary feature is shown by SR, CV, etc.

In this way, from the second segmentation means (7),
As phoneme interval information, phoneme boundary candidate information and its phoneme boundary feature information are obtained. This phoneme segment information is then supplied to the phoneme recognition means (8).

This sound B recognition means (8) uses each parameter from the acoustic analysis means (5) as a parameter for recognition processing, and executes phoneme recognition while referring to the phoneme section information from the second segmentation means (7). Recognized phoneme symbols are obtained from this sound B recognition means (8), and are supplied to the subsequent continuous speech and Daigo Confucian speech recognition means.

Although the above example is configured with hardware, the first and second segmentation means (6) and (7), the calculation part of the acoustic analysis means (5), and the phoneme recognition means (8)
can be realized by a computer.

〔Effect of the invention〕

According to this invention, feature points predicted to be phoneme boundaries are extracted from a plurality of parameters obtained through acoustic analysis, and phoneme interval candidates are determined from information on the feature points of these plurality of parameters. It becomes easier to obtain more accurate phoneme interval information. Furthermore, since the characteristics of the phoneme interval candidates are also obtained as the phoneme interval information, there is an advantage that it becomes easier to recognize the sounds fJ Kn.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of an embodiment of the present invention, and FIG. 2 is a diagram for explaining the same. (5) is an acoustic analysis means, (61) is a feature point extraction means, (
71) is a feature point integration processing means, and (72) is a phoneme boundary feature detection means.

Claims (1)

[Scope of Claims] a) Acoustic analysis means for acoustically analyzing input speech to obtain parameters for a plurality of phoneme intervals; b) Means for extracting feature points for each of the parameters; c) Features of each of the pamelata. A phoneme interval information forming device comprising means for determining phoneme interval boundary candidates and phoneme boundary features of each boundary candidate from point information, and outputting these phoneme interval boundary candidates and phoneme boundary features as phoneme interval information.