JPH04343399A - Speech feature extraction method and speech recognition device - Google Patents

Abstract

PURPOSE:To improve the accuracy of voice recognition by correcting an utterance fluctuation under noise and extracting the feature quantity of arm utterer. CONSTITUTION:This device is provided with a level measuring means 11 for measuring a level of an input signal a frequency analyzing means 12 for executing a frequency analysis of the input signal, a voice section detecting means 13 for detecting a voice section in the input signal, and a coefficient setting means 14 for setting a coefficient of the frequency analyzing means, based on a level of the input signal measured by the level measuring means 11, at the time point when the voice section is not detected by the voice section detecting means. In such a state, a level of a noise for exerting the influence on an utterer is measured, and in accordance with its level, a coefficient of a frequency analysis filter is changed, and the feature quantity of a voice which corrects a spectrum deformation of utterance by a Lombard effect is obtained.

Description

[Detailed description of the invention]

0001

TECHNICAL FIELD The present invention relates to a voice feature extraction method and a speech recognition device using the voice feature extraction method, and more particularly, to
The present invention relates to a technology for correcting speech conversion under noisy conditions, and is suitable for application to speech recognition devices used in offices, cars, factories, and homes, for example.

0002

2. Description of the Related Art It is known that the volume of speech in noise increases and the spectrum is distorted compared to speech in no-noise conditions. This effect is called the Lombard effect.
This occurs when the voice feedback to the speaker himself is blocked by the noise, resulting in the speaker speaking loudly. This is one of the reasons why speech recognition under noisy conditions is difficult. As a countermeasure against this, for example, there is a speech recognition preprocessing device described in Japanese Patent Application Laid-Open No. 3-2793, but this device has a system that allows the speaker to feed back his or her own voice at an appropriate volume. By doing so, the Lombard effect itself is prevented. However, one drawback of this method is that it requires a playback device such as a playback amplifier and headphones, and the ratio of the sound to surrounding noise (SN ratio) is limited to a certain extent, such as the inability to use a close-talking microphone. In the worst case, there is no increase in volume, so the S/N ratio becomes worse than normal speech without using this method, making recognition difficult, which is another drawback.

It is known that the formant frequency of speech subjected to the Lombard effect increases, especially in the low range. For example, according to "Study of recognition method considering vocalization deformation under noise", Onkoron, January-October 1999, 2-1-5, when there is 86 dBSPL noise, compared to when there is no noise,
The average formant frequency below 1.5 KHz is approximately 120
Hz is rising. It is difficult to perform recognition by directly using the feature values of voices with such vocalization distortions.

0004

[Purpose] The present invention has been made in view of the above-mentioned actual circumstances, and is a method for improving the SN ratio by directly using the natural efforts of the speaker to improve the SN ratio while eliminating the Lombard effect. In addition, it is possible to measure the level of noise that affects a person, and obtain voice features that correct formant frequencies shifted by the Lombard effect according to the level.Furthermore, in speech recognition in noise,
This was done with the aim of performing highly accurate recognition without being affected by the Lombard effect.

[0005]

[Structure] In order to achieve the above objects, the present invention provides (1)
a level measuring means for measuring the level of an input signal; a frequency analyzing means for performing frequency analysis of the input signal; a voice section detecting means for detecting a voice section of the input signal; a coefficient setting means for setting a coefficient of the frequency analysis means based on the level of the input signal measured by the level measurement means at a time when the frequency analysis means is not detected; , (2) from a microphone for inputting audio and an input signal obtained from the microphone, the above (1)
a feature amount extraction unit that obtains a voice feature amount using the voice feature amount extraction method described above; an input pattern generation unit that obtains an input pattern from the voice feature amount obtained by the feature amount extraction unit;
a standard pattern memory that stores standard patterns of speech registered in advance; and a recognition unit that performs recognition using the input pattern obtained by the input pattern generation unit and the standard pattern stored in the standard pattern memory and outputs the result. The present invention is characterized by a speech recognition device comprising: Hereinafter, the present invention will be explained based on examples.

FIG. 1 is a configuration diagram for explaining an embodiment of the invention as claimed in claim 1, in which a level measuring means 11 performs absolute value calculation and low-pass filtering (LPF) to measure the level of an input signal. It is calculated and output every 10ms. The frequency analysis means 12 obtains the spectrum of the input signal by 15 channels of band pass filtering (BPF), absolute value calculation, and low pass filtering (LPF), and outputs it as a voice feature every 10 ms. The level measuring means 11 and the frequency analyzing means 12 are composed of digital filters, and the center frequency of the 15-channel BPF of the frequency analyzing means 12 is set by the coefficient setting means 14. The voice section detection means 13 is the level measurement means 11.
From the time-series data of the level of the input signal obtained every 10 ms, a voice section is detected using a known two-threshold method. The speech interval detection method used here is not the essence of the present invention, and other methods may be used.

[0007] The coefficient setting means 14 includes the frequency analysis means 12
Three sets of coefficients corresponding to three center frequencies are prepared in advance for each BPF, and depending on the noise level,
Change its coefficients. The noise level is obtained by smoothing a plurality of pieces of noise level data in a section determined by the voice section detecting means 13 to be not a voice section. The two noise thresholds for distinguishing the noise level into three levels are absolute sound pressures of 70 dBSPL and 80 dBSPL, respectively, at the microphone position used (more preferably at the speaker's ear position).
It is set in advance to correspond to BSPL. This threshold value naturally varies depending on the microphone sensitivity and microphone amplifier gain of the system used. Here, the center frequency of each BPF is (a) when the noise level is 70
If equivalent to dBSPL or less, 1/3 from 250Hz
(b) The noise level is set at octave intervals, and (b) the noise level is 70~
If the noise level corresponds to 80 dBSPL, the frequency shall be the sum of each center frequency in (a) plus 50 Hz, and (c) If the noise level corresponds to 80 dBSPL or more, the frequency shall be the sum of each center frequency in (a) plus 100 Hz. do.

[0008] Therefore, the ambient noise level changes and Lo
Even if the frequency of a certain formant changes due to the mbard effect, that formant can be captured in the same channel regardless of the noise level, so the spectrum obtained by the frequency analysis means 12, that is, the feature amount of the voice, is
The influence of the Lombard effect will be corrected. In the settings (b) and (c) of this example, the center frequencies of all BPFs are raised uniformly, but the basic (a)
) settings are at equal intervals on the LOG scale, so changing settings has little effect on high frequencies, and
Since the settings of each BPF can be changed independently, there is no particular need to uniformly increase the BPF. In this example, three sets of center frequencies for each BPF are prepared and reset. However, in a system where the amount of calculation processing is not an issue, the optimum center frequency is By calculating and setting the frequency, more accurate voice features can be obtained. According to the above feature extraction method of the present invention, not only the center frequency is changed, but also the gain of some BPFs is changed.
It is also possible to change the frequency characteristics at the same time. Lomb
It is known that the ard effect involves not only a shift in formant frequency but also an increase in energy in a specific band, and the present invention can be used for this correction.
This correction method is also included within the scope of the present invention.

FIG. 2 is a diagram showing the configuration of an embodiment of the invention according to claim 2. In the figure, a microphone 1 inputs audio,
The preprocessor 2 converts the signal into an electrical signal, and amplifies the signal and cuts off the high frequency range. The A/D converter 3 converts the analog signal into one
Convert to 12-bit digital value with 6kHz sampling. The feature quantity extraction unit 10 executes the above-mentioned feature quantity extraction, and the level measurement means 11, frequency analysis means 12, voice section detection means 13, and coefficient setting means 14 are the same as in the example of FIG. be. The input pattern generation section 4 obtains an input pattern from the voice feature amount obtained by the feature amount extraction section 10 of the section determined as a voice section by the voice section detection means 13. The standard pattern memory 5 stores standard patterns of voices registered in advance. The recognition unit 6 performs recognition using the input pattern obtained by the input pattern generation unit 4 and the standard pattern stored in the standard pattern memory 5, and outputs the result. Here, the operations of the input pattern generation section 4, standard pattern memory 5, and recognition section 6 are as follows: "Word speech recognition system using binary TSP", Electrical Engineering Theory C, Vol. 108, 1983.
-October, pp. BTST known as 858-865 etc.
Although a standard voice recognition technique is used, other methods may also be used.

FIG. 3 is a diagram showing an example in which the configuration of the embodiment shown in FIG. 2 is implemented as hardware.
1 corresponds to the preprocessing section 2 in FIG. 2, and the A/D converter 22 corresponds to the A/D converting section 2. The DPS 23 includes a feature extraction unit 10, and the ROM 25 includes a standard pattern memory 5.
data, a program for the input pattern generation section 4, and a program for the recognition section 6. The CPU 27 executes the program of the input pattern generation section 4, the program of the recognition section 6, and the like. RAM 26 is used as a work memory for program execution. The I/O boat 28 is an RS
The 232C port is used to communicate with the outside, such as outputting recognition results.

[0011]

[Effect] Effect corresponding to claim 1; Since the level of noise that affects the speaker is measured and the coefficients of the frequency analysis filter are changed according to the level, Lombar
It is possible to obtain the feature amount of the voice that corrects the spectrum deformation of the voice due to the d effect. Further, since the correction is performed at the time of frequency analysis, it is not necessary to perform arithmetic processing for correction on the obtained feature amount, and recognition processing can be performed immediately from the feature amount. Effect corresponding to claim 2: Since a voice feature quantity that corrects the spectral deformation of vocalization due to the Lombard effect is obtained, accurate recognition is performed without being affected by the Lombard effect even in voice recognition in noise. be able to.

[Brief explanation of the drawing]

FIG. 1 is a configuration diagram for explaining an embodiment of the invention according to claim 1.

FIG. 2 is a configuration diagram for explaining an embodiment of the invention according to claim 2.

FIG. 3 is a diagram showing an example of a configuration in which the embodiment shown in FIG. 3 is implemented as hardware.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1... Microphone, 2... Preprocessing part, 3... A/D conversion part, 4... Input pattern generation part, 5... Standard pattern memory,
6... Recognition section, 10... Feature amount extraction section, 11... Level measurement means, 12... Frequency analysis means, 12... Voice section detection means.

Claims (2)

[Claims] 1. Level measurement means for measuring the level of an input signal; frequency analysis means for frequency analysis of the input signal; voice section detection means for detecting a voice section of the input signal; and coefficient setting means for setting the coefficients of the frequency analysis means based on the level of the input signal measured by the level measurement means at a time when no voice section is detected by the detection means. A method for extracting audio features. 2. A microphone for inputting sound; a feature extraction unit for obtaining a sound feature by the sound feature extraction method according to claim 1 from an input signal obtained from the microphone; an input pattern generation section that obtains an input pattern from the feature amount of the voice obtained in the section; a standard pattern memory that stores a standard pattern of voice registered in advance; A speech recognition device comprising: a recognition unit that performs recognition using a standard pattern stored in a standard pattern memory and outputs a result.