JPS6069700A - Voice recognition equipment - 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 <a> Technical Field of the Invention The present invention relates to a speech recognition device that recognizes input speech using phonemes as recognition units.

(b) Technical Background Speech recognition devices are classified into monosyllabic speech recognition devices, word speech recognition devices, or continuous speech recognition devices depending on the recognition unit, and are monosyllabic speech recognition devices and word speech recognition devices that target specific speakers. Devices such as speech recognition devices and small word speech recognition devices targeted at unspecified speakers have already been put into practical use.

However, a continuous speech recognition device is the most desirable means for inputting information to an electronic pointer system, and in order to put it into practical use, speech recognition devices that use phonemes as recognition units have recently been developed.

(C) Prior Art and Problems The speech recognition device that uses phonemes as recognition units has two analysis units that analyze input speech and extract features; a phoneme dictionary that stores standard features for each phoneme; and a phoneme dictionary that stores standard features for each phoneme; The main component is a phoneme recognition unit that recognizes the phoneme of the input voice by comparing the characteristics of the input voice with the standard features stored in the phoneme dictionary.

Techniques such as fast Fourier transform (FFT), filter analysis, and linear prediction method (LPG) are used in the analysis section. A method that is often used is to perform analysis at a certain analysis frame length and analysis frame period, obtain a logarithmic spectrum for each analysis frame, and compare this with standard features for each phoneme stored in a phoneme dictionary in advance.

In this way, conventionally, analysis conditions such as the analysis frame length and analysis frame period have been kept constant.

For this reason, for example, phonemes such as voiceless plosives ())-T-K) whose duration is very short compared to vowels, etc.

It has the disadvantage that it is difficult to recognize because it is easily influenced by the phonemes before and after it.

(d) Object of the Invention An object of the present invention is to improve the recognition rate in a speech recognition device that recognizes input speech using phonemes as recognition units.

(e) Structure of the Invention: 1. The speech recognition device according to the present invention includes an analysis unit that analyzes a big man's voice and extracts features, a phoneme dictionary that stores standard features for each phoneme, and input speech obtained by the analysis unit. In a speech recognition device that recognizes human speech using phonemes as recognition units, the speech recognition device includes a phoneme recognition unit that recognizes phonemes of input speech by comparing the characteristics of the above with standard features stored in the phoneme dictionary, An analysis condition setting unit is provided to set according to the recognition result in the phoneme recognition unit, and recognition is performed by changing and setting analysis conditions for input speech according to the recognition result in the phoneme recognition unit and reanalyzing. It is.

(f) Examples of the Invention The gist of the present invention will be specifically explained below using two examples.

The figure is a block diagram showing the configuration of one embodiment of the present invention.
is an AD converter that samples the input audio every 1 ms and converts it into 12-bit multi-value digital data; 2 is an AD converter;
An analysis unit 3 reads the contents of the buffer 2, analyzes the input audio for each analysis frame length, and extracts the logarithmic spectrum and pinch frequency as its characteristics.

4 is a determination unit that determines whether the speech for each analysis frame length is a voiced or unvoiced sound based on the presence or absence of a pinch frequency obtained by the analysis unit 3; 5 is a phoneme dictionary that stores a standard logarithmic power vector for each phoneme; 6 is a phoneme recognition unit that recognizes the phoneme of the input speech by comparing the logarithmic power spectrum obtained by the analysis unit 3 with the logarithmic power spectrum of the standard phoneme stored in the phoneme dictionary 5; 7 is a phoneme recognition unit that recognizes the analysis conditions in the analysis unit 3; This is an analysis condition setting section to be set.

The analysis unit 3 analyzes the input audio according to the first analysis condition described later, and then analyzes the sound! l! If recognition is not possible in the recognition unit 6, the input voice is analyzed according to either the second analysis condition or the third analysis condition described later in accordance with the judgment result of the first judgment unit 4.

In other words, in the first analysis condition, the length of one analysis frame is 12
・3 ms Also, set the analysis frame period to 6.4 ms, calculate the logarithmic power spectrum by FFT, and calculate its 1l
Calculate the pitch frequency from the cepstrum obtained as FFT.

After this, if an unrecognizable sound occurs in the phoneme recognition unit 6 and it is determined as an unvoiced sound by the determination unit 4, the analysis frame length is set to 6.4 meters and the analysis frame period is set to 3.2 m3 as the second analysis condition. death.

Obtain the logarithmic power spectrum by FFT, and inverse FF
Calculate the pitch frequency from the cepstrum obtained as T. Compared to the first analysis condition, the second analysis condition has a shorter analysis frame length and analysis frame period, so it is less susceptible to the influence of adjacent phonemes. The recognition rate can be improved.

In addition, if the phoneme recognition unit 6 fails to recognize the phoneme, and the determination unit 4
If it is determined to be a voiced sound by Increase the pitch frequency. The third analysis condition has a longer analysis frame length than the first analysis condition, and is therefore suitable for recognizing vowels with a long duration, and particularly improves recognition accuracy by shortening the two-analysis frame period. be able to.

(g) Effects of the Invention As described above, the present invention has the effect of improving the recognition rate in a speech recognition device that recognizes input speech using phonemes as recognition units.

[Brief explanation of drawings]

The figure is a block diagram showing the configuration of an embodiment of the present invention. In the figure, 3 is an analysis unit, 5 is a phoneme dictionary, and 6 is a phoneme recognition unit 27.
is an analysis condition setting section.

Claims (1)

[Claims] an analysis unit that analyzes input speech and extracts features; a phoneme dictionary that stores standard features for each phoneme; and a phoneme dictionary that stores standard features for each phoneme; and Equipped with a sound B recognition unit that recognizes phonemes,
In a speech recognition device that recognizes input speech using phonemes as recognition units, an analysis condition setting unit is provided that sets analysis conditions in the analysis unit according to the recognition results in the phoneme recognition unit, and A speech recognition device characterized in that the analysis conditions of input speech are changed and set, and the analysis is re-analyzed.