JPS6043697A - Consonant and vowel boundary detection device - 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 Field of the Invention The present invention relates to a boundary detection device for detecting boundaries between consonants and vowels in monosyllable recognition.

Configuration of the conventional example and its problems The processing configuration for detecting the boundary between consonants and vowels in the conventional monosyllable recognition device is to sequentially calculate the interframe distance D=dlS(xt i
+xtl-1-1) The frame in which the distance between 7 +/- is equal to or greater than a certain threshold was defined as the boundary between a consonant and a vowel. However, with this configuration, although the processing time is slightly faster, the inter-frame distance is susceptible to fluctuations due to the presence of unsteady transitions near the boundaries between consonants and vowels, which poses a major problem in detection accuracy. This was the cause of a significant decrease in the recognition rate. Therefore, in a method that slightly improves the detection accuracy, after completing the detection of the speech sections at the beginning and end of a word, the normalized correlation coefficient between the stationary part of the speech section and the vowel standard pattern is calculated for each frame, and a fixed threshold value is calculated. The average vector of the stationary part is calculated, and the normalized correlation coefficient is sequentially calculated in the direction of the beginning of the word.The frame where this value is below a certain threshold is determined as the boundary between a consonant and a vowel. there was. However, even in this case, since the consonant part of a monosyllable has a large influence on the vowel part 1, the fricative /s/c/h/ and the plosive /p/l/ are affected by / and so on. /b/d/q/ more nasal consonants//
rr1//r1//q/ has problems in practicality, such as variations in the threshold value of the normalized correlation coefficient and influence of the speaker.

Purpose of the Invention The present invention solves the above-mentioned conventional problems.The present invention is aimed at solving the above-mentioned problems of the conventional art. Calculate the normalized correlation coefficient only for the upper channels among the channels that give It is an object of the present invention to provide a consonant-vowel boundary detection device that can stably improve consonant-vowel boundary detection accuracy and enable real-time processing.

Structure of the Invention The present invention detects the beginning of a word based on the power sequence of input speech,
Vowels (A/i/U/E 10/X) which are standard patterns memorized in advance for each frame (xtl)
Calculate the distance to the frame, and if there are consecutive frames for which the distance is smaller than a certain threshold θ7 for a certain period or more, the frame period [1sss to 1. Let essl be the vowel stationary section of the input voice, calculate the center of the vowel stationary section) (N: average number of frames, j: center frame of the vowel stationary section of the input voice), and let M be the number of channels of the average vector Xav. Then, the upper L channel (L
<M), and for these L channels,
The average vector xav and the frame m before the stationary part (m
<i, 8B) Normalized correlation coefficient C0 The threshold value θ that the normalized correlation coefficient C0 is stored in advance. . 1(
θ. . 1〈1) Frame 1゜ is less than or equal to 1〈1〉. 1 is detected,
Furthermore, a threshold value θ is stored in advance. . 2(θ..2〈θ.

. 1〈1) or less, frame 1゜02'' is detected, the center frame 1COICO-''COI+1CO21) of the frame 1゜.1 and frame 1c02 is detected as the boundary between a consonant and a vowel, and the end of the word is detected using the power sequence and processed. The boundary between consonants and vowels can be detected by absorbing fluctuations in the transition between consonants and vowels and fluctuations in the transition between vowels depending on the type of consonant, as well as fluctuations in the constant vowel interval depending on the speaker. The aim is to stably improve accuracy, shorten processing time, improve the recognition rate of monosyllabic speech recognition, achieve real-time processing, and put it into practical use.

DESCRIPTION OF THE EMBODIMENTS FIG. 1 shows a block diagram of a consonant-vowel boundary detection device according to an embodiment of the present invention.

An input section fi, an A/D converter 1, a voice section detection means 2 that detects the beginning and end of a word of an input voice signal, for example, by using a certain threshold value to detect a change in the power series, and a filter that detects, for example, a change in the power series at a certain time interval from the voice time series.・Bank output series or 1d
It consists of a feature series conversion means 3 that converts into a feature series (Xti) such as an output series of LPc coefficients, and a feature series storage unit 4 that stores the feature series (Xtil) over a certain interval. 5 is a feature vector (Xti) of a large sword series. The distance from the vowel standard pattern 5'' is calculated by the distance calculation means 5', and the section where the distance is equal to or less than a certain threshold value is defined as a constant vowel section.

Reference numeral 6 denotes a vowel stationary section average vector calculating means for calculating a vowel stationary section average vector (xav) 6' from several frames near the center frame of the vowel stationary section detected by the stationary section detecting means 5. 7 is the vowel stationary part flat vector (Xa
This is an average vector peak channel detection means that detects L channels having peak values from all channels of v16'. The boundary detection unit between a consonant and a vowel is a feature series storage unit 4.
The feature vector before the vowel stationary section (Xtm
The normalization coefficient calculation means 8 calculates the normalized correlation coefficient between l and the vowel stationary part normal vector (Xav) 6' only in the detected channel by the average vector peak channel detection means 7, and the normalization correlation coefficient Threshold value θCo19″ with a constant number
and threshold θ. . Frame l that is less than or equal to g///. . 1゜1
co2, the arithmetic mean of the frames 1co1 and 1co2 ((1co1''c
o2 V2) and outputs the boundary frame signal 9 by dividing it into a boundary frame signal 9.
1, then outputs an instruction (not shown) to store the vector sequence (Xt) 3' from the feature sequence conversion means 3 in the feature sequence storage section 4, while The signals of the start 15 and the end of the stationary section 16 are input, the stationary frame instruction 13 is outputted to the feature sequence storage section 4, and the vowel stationary section average vector calculation means 6 calculates the vowel stationary section average vector tXav)6''&. The average vector peak channel detection means 7 detects the peak channel of the vowel stationary part average vector (xav) 6', and for this L channel, the vowel stationary average vector (
xav) 6' and the feature vector before the start of the steady interval 15 (
xtml by the comparison sirem instruction 14, the normalized correlation coefficient is calculated by the normalized correlation coefficient calculation means 8, and the boundary frame signal 9'' detected by the boundary frame detection means 9 is input, and the consonant/vowel boundary frame 17 is inputted. It is a comprehensive control means that outputs.

The operation of the boundary detection accuracy of this embodiment configured as described above will be explained below using FIG. 2. In the figure, the original waveform of the input monosyllabic voice (C+V) is shown at 18. The vector sequence [X, +3' as the output of the feature sequence conversion means 3 is shown in 3''.The feature vector 3'' of the input signal includes the beginning of the word's initial voice 11 and the voice end of the word's end 12.
, and a vowel steady section sandwiched between the steady section start 15 and the steady section end 16 detected by the steady section detecting means 5 is shown at 19. A vowel stationary average vector (Xa
v) e' is shown as 6". The L peak channels detected by the average vector peak channel detection means 7 are also shown by Chromeno-Sochi.

Figure 20 shows how the normalized correlation coefficients are calculated in L channels between the vowel stationary average vector (, Xa, 16' and frames before the start of the stationary interval. Condition, threshold value θ 9″, and threshold value θ.

Boundary frame signal 01 by g/// 9" is shown in 21.

Input speech is converted into a digital sequence by the A/D converter 1, and converted into a characteristic sequence 3' at regular time intervals by the characteristic sequence converting means 3. Further, when the voice section detection means 2 detects the voice start 11 from the power series,
The feature series storage unit 4 starts storing feature series for each predetermined section 3''.

At the same time, the distance from the vowel standard pattern 5'' of each frame is calculated by the steady-state detection means 5, and the distance reduction means 5' calculates the distance between each frame and the vowel standard pattern 5'', and the intervals where the distance is equal to or less than a certain threshold are determined as vowel steady-state intervals, and the steady-state interval start 15 and the steady-state end end. The initial constant part average vector calculation means 6 calculates the vowel steady part average vector (xav) 6' for the feature vectors in the feature sequence storage unit 4 of the vowel stationary section, and calculates the vowel steady part average vector (xav) 6'. Only for the L peak channels of the vowel stationary part average vector (xav) 6' detected in step 7, the normalization of the feature vector (xt: i < m) of the frame before the start of the stationary interval 15 and the vowel stationary part average vector The normalized correlation coefficient calculation means 8 calculates the normalized correlation coefficient 20, 21. A certain threshold value θ..19″ or less and θ. . g/7/below is the first frame 'coj
+1co2 is detected, its arithmetic mean is detected as boundary frame signal 9'', and consonant-vowel boundary frame output 1
Set it to 7. The end of the word is detected by the voice section detecting means 2 in parallel with the above processing, and the end of the voice is regarded as 12, and the next utterance is in a standby state. This parallel processing enables real-time processing.

By adopting the above configuration, fluctuations in the transition part at the boundary between a consonant and a vowel, fluctuations in the vowel transition part depending on the type of consonant, and fluctuations in speakers can be absorbed, and the accuracy of detecting the boundary between a consonant and a vowel can be stabilized. Furthermore, real-time processing can be achieved.

FIGS. 3 to 8 show examples in which the above-mentioned embodiment apparatus is actually applied to syllables of typical consonant types in the O stage. In each figure, the horizontal axis is the duration from the beginning of the word (
ms) is shown at 9.1 frame 6m5eC, the first black line is the beginning of the word, the second black line is the segment boundary, and the third black line is the beginning of the word.
The th black line indicates the vowel stationary interval center. Also each figure a
shows the temporal change in the power of monosyllabic speech, and the same figure shows the temporal change in a similar feature vector series (here, the frequency running spectrum). It can be seen that the boundaries between consonants and vowels are well detected in all monosyllables.

Figures 3 and 4 show typical plosive consonants /ni/ and /G/, Figure 5 shows typical fricative consonants /S, Figure 6 shows typical nasal consonants /N/, and Figure 7 shows typical Target sound /R/.

Figure 8 shows typical aspirators/HA. From these facts, it can be seen that it is not affected by variations caused by any consonant type.

Effects of the Invention The present invention performs consonant and vowel detection processing by detecting the beginning of a word using a power sequence, and after detecting the beginning of the word, calculates the distance from the standard vowel turn in the feature vector series, and detects frames whose values are below a certain threshold. A continuous section of a certain length or more (isss
~1e881 is the vowel stationary section, and the vowel stationary section average vector (Xav) of several frames in the center of the vowel stationary section is calculated,
This vector h.

- Detect L channels having a
v) and the frame m before the vowel stationary interval (m < i,
Normalized correlation coefficient C0θco1 with 881 (<1) or less and threshold value θ. . The arithmetic mean of the first frames below 2 (〈θ..1 × 1) is used as the boundary frame between consonants and vowels, and the detection of word endings is processed in parallel to prepare for the next utterance. Night envoy! :1! Separation [leaping thing'
In summary, by adopting the above configuration, it is possible to absorb fluctuations in the transition part at the boundary between a consonant and a vowel, as well as fluctuations in the vowel transition part caused by the type of consonant, and also to absorb fluctuations in the vowel steady interval due to speaker variations. By absorbing this information, it is possible to stably improve the accuracy of detecting the boundary between consonants and vowels, and furthermore, real-time processing is also possible, thereby achieving practical use.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of a consonant-vowel boundary detection device according to an embodiment of the present invention, FIG. 2 is an explanatory diagram for explaining the operation of the embodiment device shown in FIG. 1, and FIGS. Figure 8 shows monosyllables A■/Go/, /So/, /No/ with consonants of the typical consonant classification of the upper syllable in the same example device.
, /RO/, /HO/ are characteristic diagrams showing the implementation results when inputting 01... A/D converter, 2...
...Voice section detection means, 3...Feature sequence conversion means, 4...Feature sequence storage section, 5'...
・Distance calculation means, 5″...Vowel standard Bakun,
5...Steady part detection means, 6...Vowel steady section average vector calculation means, 7...Mean vector peak detection means, 8...Normal ratio phase °Relational number word"thin means, 9... Boundary frame detection means.

Claims (1)

[Claims] The input speech is expressed as a time series pattern of feature vector xti (Xt
l, xt2. . . . IxtN), a speech section detection means that extracts a speech section from the input signal, and a frame whose distance from the vowel standard pattern is less than a certain threshold in the feature vector series. a stationary part detecting means for detecting a section where the vowel stationary section continues for a certain section length or more as a vowel stationary section; and a normalized correlation coefficient divider for calculating a normalized correlation coefficient between the average vector of the vowel stationary section and frames before the stationary section. calculating means, sequentially from the stationary interval toward the beginning of the word, calculate the normalization phase for the plurality of upper channels among the channels having the maximum value of the vowel stationary interval average vector (Xtil) and the channel having the minimum value. A consonant-vowel boundary detection device, characterized in that it compares whether a value obtained by a relational coefficient calculation means is below a certain threshold value, and sets the frame in which the value is below the threshold value for the first time as a consonant-vowel boundary frame. .