JPS61118800A - Voice analyzer - 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] [Industrial Application Field] The present invention relates to a monosyllabic speech recognition device, which is used to accurately recognize unvoiced plosives among monosyllables, and to detect the point at which they plosive. It relates to a speech analysis device.

Speech recognition devices are classified by recognition units into monosyllabic speech recognition devices that recognize syllables as units, word speech recognition devices that recognize words as units, continuous speech recognition devices that recognize continuous speech, and the like.

Among these, monosyllabic speech recognition devices are devices that input speech to be recognized syllable by syllable and, for example, use the following procedure to determine whether the speech is classified into one of approximately 60 types of syllables in the case of Japanese. It is.

i) A microphone converts the voice uttered by a speaker into an electrical signal for each single syllable.

ii) 9. For example, 10.
It samples every kHz and performs AD conversion.

1ii) The AD-converted audio signal is analyzed every 10 m5 using an analysis window with a length of, for example, 20 m5, and features such as a power spectrum are extracted to create a feature time series for, for example, 40 windows.

iv) Compare the obtained feature time series with a standard feature time series prepared in advance for each type of monosyllable to determine the degree of similarity.

(2) Determine the type of syllable to be recognized based on the obtained similarity.

FIG. 2 is an example of the sound and voice signal for the voiceless plosive 1 ba (pa), and ■ is the area that characterizes the consonant p.

■ is the region that characterizes the vowel a, and ■ is the region that transitions from the consonant p to the vowel a, i.e., r, gj, j. Compared to the vowel region ■, the consonant region I and transition region ■, which most characterize voiceless plosives, are extremely This is a rapid phenomenon.

Therefore, in recognizing voiceless plosives, it is very important to accurately extract the characteristics of the consonant region and transition.

[Conventional technology]

FIG. 3 is a block diagram showing the configuration of a conventional example of a monosyllabic speech recognition device, in which 1 is a microphone that converts monosyllabic speech uttered by a speaker into an electrical signal; If the audio signal is 1, for example 10k
AD converter that samples every Hz and performs AD conversion, 3
4 is a buffer that temporarily stores the output of the AD converter 2, and 4 is a buffer that analyzes the audio signal stored in the buffer 3 at a cycle of 10 m5 using an analysis window of +20113 length, extracts features such as power spectrum, A feature extractor 5 creates a feature time series, a feature dictionary 5 registers standard feature time series prepared in advance for each type of monosyllable, and a feature 6 registers the feature time series obtained by the feature extractor 4 in the feature dictionary 5. 7 is a determining unit that determines the type of syllable to be recognized based on the similarity obtained by the matching unit 6.

[Problem that the invention seeks to solve]

In the above configuration, the feature extraction unit 4.

As shown in Fig. 2, an analysis window of 9, for example, 20 S length is
■-・・■-・・1 While moving at a period of, for example, 10 m5, a feature time series is created by extracting the features in each section, but 2 There are no restrictions on the starting point of the analysis window. Ta.

For this reason, the position of the analysis window over the "consonant region" becomes irregular, and the obtained features differ depending on the case of ■ and the case of ■°. Therefore, the characteristics of consonant region I can be stably and accurately extracted. The problem was that it couldn't be done.

[Means for solving problems]

A voice analysis device according to the present invention is a means for solving the above problems, and includes a first arithmetic circuit that generates a time series of voice power from a voice signal, and a time series of zero crossings per unit time from the voice signal. A second arithmetic circuit that generates a sequence, and a detection circuit that detects the point of rupture of a voiceless plosive in the audio signal from the output of the first arithmetic circuit and the output of the second arithmetic circuit. It is.

[Effect]

FIG. 4 is a diagram showing the relationship between the voice signal A of the voiceless plosive "pa", its logarithmized voice power change B, and the number of zero crossings change C. The voice signal A appears from around point P. Therefore, along with the audio signal A, the audio power B is also increasing.

In particular, the change in voice power from the point of rupture to the rise of the vowel is large. Also, if we look at this in terms of the number of zero crossings C,
It shows a fairly rapid decrease in the amount of steel at the point of rupture, and the value is small compared to that in the silent section and vowel part.

The present invention has been made with attention to such properties, and the point of rupture of a voiceless plosive is detected as the intersection point Q between the change B in voice power and the change C in the number of zero crossings. It is something.

〔Example〕

The gist of the present invention will be specifically explained below with reference to an embodiment shown in FIG.

FIG. 1(a) is a block diagram showing the configuration of an embodiment of the present invention, in which 1' is a microphone that converts monosyllabic speech uttered by a speaker into an electrical signal, and 21 is a microphone that converts monosyllabic speech uttered by a speaker into an electrical signal. 9. For example, 1Qk
AD converter that samples every Hz and converts it into D, 3
° is a buffer that temporarily stores the output of the AO converter 2°.

8 is the first arithmetic circuit that generates a time series of audio power from the audio signal, which is converted into an electrical signal by a microphone 1°, sampled every 10 kHz, AD converted, and 9 time series X (represented as 111). Buffer 3”
A monosyllabic speech signal stored in , using an analysis window with a length of 200 samples, that is, a length of 20 m5.

With a cycle of 10m5.

P (1) = log (Σsq, (x(nl) )
---- Calculate (1) (where Σ represents the sum of n = 1 to 200) and logarithmized power time series CP (11)
The output is stored in the buffer 9.

10 is a second arithmetic circuit arithmetic circuit that generates a time series of the number of zero crossings per unit time from an audio signal;
Using an analysis window of 200 samples length, the monosyllabic speech signal X (nl stored in
) and stores the output in the buffer 11.

12 and 13 are normalization circuits, which respectively calculate the time series (P(j)) of the logarithmized audio power stored in the buffer 9 and the time series (Z (N)) of the number of zero crossings stored in the buffer 11. ), the maximum value is 1
Multiply the normalization coefficient Kl and by 2 so that

P (1) ' -Kl x P (1-------
(2) Z (1) ' - to 2 x Z (n)
------- (3) Output a time series of normalized logarithmic voice power (P(6)'') and a time series of the number of zero crossings (Z(j)').

14 is a detection circuit that detects the plosive point of a voiceless plosive in a monosyllabic speech signal from the output of the first arithmetic circuit 8 and the output of the second arithmetic circuit 10, which is output from the normalization circuit 12. Time series of normalized logarithmic speech power (p(It
)'), and the time series of the normalized number of zero crossings output from the normalization circuit 13 (Z(jri'))
As point Q) in the figure.

Detect the point of rupture.

FIG. 1(b) shows a specific circuit configuration example of the bursting point detection circuit, in which 15 is a subtraction circuit, 16 is a two-digit shift register,
17i code identification circuit.

For subtraction! 15 calculates the difference between the normalized number of zero crossings Z(N) output from the normalization circuit 13, the normalized logarithmic power output from the normalization circuit 12, and PCl>', and sends the output to the shift register 16. It will be done.

The code identification circuit 17 calculates the product of the stored contents of each digit of the shift register 16, and generates an output signal when the sign changes from positive to negative.

〔Effect of the invention〕

As explained above, according to the present invention, the point of rupture of a voiceless plosive can be stably detected, so the analysis window used for feature extraction in a monosyllabic speech recognition device is synchronized with the detected point of plosive. As a result, the characteristics of the consonant region in voiceless plosives can be extracted stably and accurately, thereby improving the accuracy of the monosyllabic speech recognition device.

[Brief explanation of drawings]

FIG. 1(a) is a block diagram of one embodiment of the present invention. FIG. 1 (bl is a diagram showing a partial configuration example of the same embodiment. FIG. 2 is an explanatory diagram regarding industrial application fields. FIG. 3 is a block diagram of the prior art. FIG. 4 is an explanatory diagram regarding the action. In the figure. 8 is the first arithmetic circuit, 10 is the second arithmetic circuit. ¥-3t3 I+WJ

Claims (1)

[Claims] a first arithmetic circuit that generates a time series of audio power from an audio signal; a second arithmetic circuit that generates a time series of the number of zero crossings per unit time from the audio signal; and an output of the first arithmetic circuit. A speech analysis device comprising: a detection circuit that detects the point at which a voiceless plosive in the speech signal ruptures from the output of the second arithmetic circuit.