JPS6163899A - Monosyllabic 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 (Technical Field of the Invention) This invention analyzes the frequency of an input voice (single sound m) using a group of band-pass filters, and generates an output pattern by dividing it into a consonant part and a vowel part; The present invention relates to a monosyllabic speech recognition device that recognizes input speech by matching it with a standard pattern that is registered in advance.

(Technical background of the invention and its problems) Conventionally, in the field of monosyllabic speech recognition, input speech (funeral sound W
J) is frequency-analyzed to calculate the spectral envelope line, and the frequency band with concentrated energy obtained by analyzing this envelope line, that is, the formant frequency, is extracted as a characteristic parameter of the input speech (monosyllabic), and this formant frequency A monosyllabic speech recognition device has been proposed that performs matching processing by comparing temporal changes (speech patterns) with standard patterns. In such conventional monosyllabic speech recognition devices,
As shown in FIG. 1, an audio signal As input from a microphone 1 is amplified to a predetermined amplitude level by an amplifier 2, and its output is input to bandpass filters 31 to 3n each having a different bandwidth. The outputs of these bandpass filters 31 to 3n are input to envelope detection circuits 51 to 5n via rectifiers 41 to 4n, respectively, and the envelope outputs E1x from the envelope detection circuits 51 to 5n are input to envelope detection circuits 51 to 5n.
En is input to the multiplexer 6 and also to the averaging circuit 7, and is sequentially switched and output together with the envelope outputs El-En of the envelope detection circuits 51 to 5n. The output of the multiplexer 7 is digitized by an AD converter 8 and then stored in a memory 9.

Here, the output waveform a (, EM) of the averaging circuit 7 is divided into three parts, as shown in FIG.
Conventionally, as shown in Fig. 83, the output waveform a of the averaging circuit 7 is observed, and this waveform a corresponds to a preset energy level
When a value larger than 0 is continuously shown for a predetermined time period of 71 g1 or more, it is determined that a voice has been input.

Then, a standard pattern for registration is stored in advance at a predetermined location in the memory 9, and a reference pattern 6 is used to extract the input voice intensity newly written to the memory 9 and the consonant part stored in advance in the memory 9. are moved in parallel along the time axis as shown by arrows P and Q in FIG.
The position (point in time) where the input voice intensity a most closely matches the input audio intensity a was determined.

Here, when the reference pattern b and the input speech intensity a match, the time 72 portion shown in FIG. 3 is extracted as a consonant pattern, and the maximum value EP of the input speech intensity a is multiplied by a constant (for example, 0.9). The position (time point) at which the value Ema intersects with the input speech intensity a is adopted as a vowel pattern. These extracted consonant patterns and vowel patterns are compared with consonant patterns and vowel patterns of standard speech previously stored in the memory 9, and the most similar pattern is selected and output.

However, such conventional monosyllable recognition processing has the disadvantage that when similar monosyllables such as "shi" and "chi" or "su°' and ゛tsupa" are input, the consonant part cannot be detected accurately. Furthermore, as shown in Fig. 5, when the maximum point (peak) of the input speech intensity a overlaps with the consonant data portion and appears in the average output of the bandpass filter group, the identification of the input speech intensity a and the reference pattern is Figure 6 (A) or (
As shown in the shaded area d in B), there was a problem in that the consonant part could not be extracted accurately.

(Objective of the Invention) An object of the present invention is to provide a monosyllabic speech recognition device that does not have the above-mentioned drawbacks and problems.

(Summary of the Invention) This invention analyzes the frequency of input speech using a group of band-pass filters, and calculates the distance between an output pattern generated by dividing it into consonant parts and vowel parts, and a standard pattern registered in advance. This relates to a monosyllabic speech recognition device that recognizes input speech as monosyllables.It extracts the formant for each input speech, selects a bandpass filter corresponding to the formant, and generates speech intensity from the output. The consonant part and vowel part of the input speech are detected, and the output pattern of the band-pass filter group is compared with a standard pattern to identify the consonant part and vowel part of the input speech.

(Embodiment of the Invention) As shown in FIG. 7 corresponding to FIG.
3n, and also recognizes the input speech (single sound! Concerning syllable recognition processing.・Here, the part X shown in Figure 5 corresponds to the low frequency band (e.g. 200 to 500) as shown in Figure 8.
)1z) has a strong spectrum.
Bar) BS (e.g. “Ba”, “Da”, etc.
) appears, or as shown in Fig. 9, multiple spectra sp appear in a certain frequency band (for example, 200 to 2000 Hz) (hereinafter referred to as ``nasal-eurmur'') (for example, ``ma°'', ``na'', . . . ) appears, resulting in the input speech intensity of X shown in FIG. 5, which affects the extraction of consonant parts. Note that in FIGS. 8 and 9, TS indicates a time spectrum pattern. Furthermore, the detection of RP at the boundary between the consonant part and the vowel part (hereinafter referred to as the reference point), which affects the consonant recognition rate,
Identification based on the input speech intensity using the output of all bandpass filters in the conventional method is difficult, resulting in large discrepancies and a decline in the recognition rate due to the effect of time. Therefore, input using only formant information without the above-mentioned problems is recommended. Generation of audio intensity is required. This invention was made with this in mind.

The invention will now be explained with reference to FIG.

The input voice is frequency-analyzed by band-pass filter groups 31 to 3n, and a certain frequency band (for example, 200 to 5
The frequency band outputs E5 to En excluding 00FIz) are taken into the comparator IO from the time when it is determined that the voice has been input (vowel stable part), and the number of channels with the strongest level (for example, 5) is selected. Then, the signal is added to the counter 11. Then, the counter 13 clocks at a predetermined frequency.

For example, it is determined whether or not 150 samples of data have been captured by counting the Kuvarus CP, and when the data of 150 samples has been captured, the channel selection circuit 12 selects 5 channels from the largest value based on the contents of the counter 11. Select and output the data. At the same time, 70 pieces of audio data stored in the memory 9 are read out to the addition/average circuit 14, and this and the data of the channel selected by the channel selection circuit 12 are added and averaged every time by the addition/average circuit 14. This averaged data is further passed through a low pass filter (LPF) 15 for the purpose of removing high frequency components,
The input audio intensity ENV obtained thereby is sent to the multiplexer 6. Furthermore, the consonant part and vowel part of the input speech (monosyllabic) are detected from the input speech intensity ENV, and
The output patterns of the bandpass filter groups 31 to 3n are compared with standard speech patterns stored in advance in the memory 9 to identify consonant parts and vowel parts of the input speech.

(Effect of the invention) By using the above method, FIGS.
) Extract the reference point RP accurately as shown in 1. In addition to improving the recognition rate of consonants, it also absorbs the influence of lateral differences,
It has advantages such as high stability of the standard pattern. That is, FIG. 1θ (A) shows the conventional input voice intensity,
FIG. 2B shows the input voice intensity according to the present invention, and it is clear that pattern matching can be performed smoothly. Further, since the above-mentioned hardware configuration can be easily programmed using computer software, there is an advantage that high-performance speech recognition can be performed using only an inexpensive system configuration. Furthermore, the voice recognition device of the present invention can be easily applied to input means such as typewriters and computer phototypesetting, or to the operation and control of a single machine or device.

[Brief explanation of the drawing]

Fig. 1 is a block diagram showing an example of a conventional monosyllabic speech recognition device, Fig. 2 is a time chart for explaining the temporal transition of speech energy, and Figs. 3 and 4 each illustrate the speech matching process. Time chart to explain,
5 and 6 (A) and (8) are time charts for explaining the conventional monosyllabic speech recognition process, FIG. 7 is a block diagram showing an embodiment of the present invention, and FIGS. 9
The figure shows the improvement items of the conventional device, Figure 1O (A), (
B) is a diagram showing the improvement content of the present invention. l...Microphone, 2...Amplifier, 31-3n.
... Band pass filter, 41~4n... Rectifier, 5
1 to 5n...Envelope detection circuit, 6...Multiplexer, 7...Envelope generation circuit, 8...AΩ converter, 9
...Memory, lO...Comparator, 11.13...Counter, 12...Channel selection circuit, 14...Addition/averaging circuit, 15...Low pass filter. Applicant's agent Yu Angata No. 3 Figure 22 Figure 32 L 4 Figure 51 6 Hard (A) <s) Figure 6 Figure 9 Figure lθ Figure 13 Written amendment to the procedure (formality) 1985 February 20th

Claims (1)

[Claims] By frequency-analyzing the input voice using a group of band-pass filters and calculating the distance between the output pattern generated by dividing the input voice into consonant and vowel parts and a pre-registered standard pattern, the input voice can be converted into monosyllables. In the monosyllabic speech recognition device, a formant is extracted for each input speech, a bandpass filter corresponding to the formant is selected, a speech intensity is generated from the output, and the consonant and consonant portions of the input speech are extracted. Along with detecting the vowel part,
A monosyllabic speech recognition device characterized in that a consonant part and a vowel part of the input speech are identified by comparing the output pattern of the group of bandpass filters and a standard pattern.