JPS59176799A - Monosyllabic voice recollation system - 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] The present invention provides a speech recognition device in which the initially selected candidates are /, such as "da" and rzaJ, rgaJ and "ka J".
The present invention relates to a monosyllabic speech re-verification method that allows for highly accurate identification in the cases of d/ and /Z/ and /g/ and /ni/.

[Conventional technology and problems]

When performing speech recognition, as shown in Figure 1, the unknown speech input Z is frequency-analyzed by appropriate means such as FFT or a bandpass filter, these parameters are extracted, the characteristics of the unknown speech input are determined, and this is stored in a dictionary. DP comparison is performed with all stored known voice feature parameters, and those with similar features are output as single or multiple recognition results. In this case, when performing DP verification,
DP matching is performed in the same way for any syllable. However, when performing monosyllable DP matching, /ka/ and /g
a /, /sa/ and /ta /, /pa/ and /la/
Because there are words such as / / ma / and / na / that have similar characteristics, it is not always possible to completely recognize monosyllabic speech using only the DP method. Sometimes. Therefore, in such a case, it is necessary to carry out re-verification 2 according to the initially extracted candidates to perform accurate recognition.

[Purpose of the invention]

The object of the present invention is that the candidates extracted by the first matching are "da" and rza' J, rgaJ and "ka J," /d/ and /z /, /g/ and /ni/. It is an object of the present invention to provide a monosyllabic speech re-verification method that can accurately identify monosyllabic speech in certain cases.

[Structure of the invention]

In order to achieve this objective, the monosyllabic speech re-matching method of the present invention selects a plurality of re-matching candidates 2 obtained by matching pre-registered monosyllabic speech and unknown input monosyllabic speech, and In a speech recognition device that performs re-matching using an identification matching method suitable for matching candidates, it is equipped with a medium-high pass filter and a low-pass filter 2, and each filter output has a threshold value that changes depending on other filter outputs and the point in time when the provisional threshold value 2 is exceeded. death,
a starting point detection means for determining the point at which the mid-high band filter output and the low band filter output exceed their respective thresholds as the starting point of the voiced consonant part and the starting point of the angular consonant part, respectively, and DI) matching means 2;
The present invention is characterized in that at least one of the starting points is used as a starting point for Dr inquiry based on a combination of re-verification candidates.

[Summary of the invention]

Prior to describing the invention in detail based on embodiments, a brief explanation will be given with reference to FIG. 2. (A) to (C) in FIG. 2 show the audio energy waveforms of "ka J", and (2) to (e) show the audio energy waveforms of "ga". And (a)
, (b) are the original voice energy waveforms of "ka" and rgaJ, (b) and (e) are "ka" and "ga"
The energy waveform 2 of the waveform passed through the low-pass filter of `` is shown, and (c) and (f) are the energy waveforms of the waveform of the waveform passed through the middle-pass filter 7 of [ka J and rgaJ. The voiced starting point T obtained by the low-pass filter as shown in
A slightly weaker part from 1 to the silent start point to the vicinity of the buzz bar (
buzzbar), but in /g/ and /Z/, this part of the buzzbar may fluctuate, such as being larger or smaller. Therefore, excluding this part, for example, for "ka J,"
DP verification is performed from the voiced starting point T1 and unvoiced starting point T4 for ga J, so a part of the buzzer may appear or not appear, so DP is compared from the unvoiced starting point T4.
Verify 7 ko, or also "ka" is r ga
j is also the later of their respective starting points (, rkaJ
The DP light is performed starting from the voice start point T3 (in RGAJ) and the silent start point T4 (in rgaJ). As a result of this, it is possible to perform the DI verification again using the starting point query criterion where 1 energy exists, and it has been found that highly accurate re-inquiry results can be obtained. In particular, very good results were obtained when DP verification was performed using the later of each starting point as the starting point.

Example One embodiment of the present invention will be explained with reference to FIG.

FIG. 3 is a configuration diagram of an embodiment of the present invention.

In FIG. 3, l is a low-pass filter, for example 5C
Those that ignore frequency components from IHz to 350Hz,
2 is a mid-high pass filter, for example IJG (z to 4
．． 9. Those that pass the frequency components of Nlz, 3゜4 are respectively back-calculation units that calculate the energy for each extracted frequency component, 5.6 are each threshold determination units, and 7 and 8 are each The starting point is 3μ, and the one that detects the tentative end τ, which is uninvented, is χ iron. Reference numeral 9 denotes a selector which selects and outputs a voiced start (voiced start or unvoiced start), and the selection of the output desired is made, for example, by writing in advance in the cheagle 10 and checking it. Reference numeral 11 is a DI) collation section, which performs re-verification. Of course, this Dr headlight 811 used for the first inquiry may be used again.

First, in FIG. 3, the voiceless starting point tuo and the voiced starting point tv
o Explain how to obtain accurate information.

Power calculation unit 3 for the low frequency components that have passed through filter 1
In addition, the energy PwL is calculated in the power calculation section 4 for the middle and high frequency components that have passed through the filter 2.

The voiced start end detection unit 7 basically uses the energy PwL
When tV' exceeds a predetermined threshold, the tentative beginning tV' of a voiced consonant is extracted. However, even if the original energy PwL is sufficient, leakage occurs when the mid-high frequency component energy PyH is large. Crabs grow bigger. For this purpose, the threshold value determination unit 5 is configured to introduce the value of the energy PwFL and determine the threshold value. Also, the threshold value determination unit 6
The same is true for , and the threshold value is determined by introducing the values of the energies P to vL.

The extraction of the temporary candy end tsu' in the start end detection units 7 and 8 is performed as follows.

[■] Processing in the silent start end detection unit 8.

An observation window having a time width of 10 ms is scanned for the above energy PwH so that samples are scattered at intervals of 2 ms. The energy within the observation window at the time position becomes the threshold THu = 3.0 + 0. IXPwL-
(When over 11, the temporary beginning tti for voiceless consonants
'v Extract.

Processing performed by the CIO start end detection unit 7.

For the above energy PwL, an observation window with a time width of 5 ms is scanned so that samples are scattered at intervals of 2 ms (.Then, the energy within the observation window at the time position of (1) is the threshold value THL1= 0.5 - (2) or (11) the temporary starting point tv obtained by the equation (2)
If ' is later than the tentative starting point tu', the energy within the observation window at the above time position is the threshold value THLz = 1
0.0 + 0.5 X PwH(3)? The tentative beginning of a voiced consonant with either "TV"
a' Extract.

The temporary starting points tu' and tv' are obtained as described above, but the starting points tu and tv are extracted by performing the following processing in the vicinity of these temporary starting points.

(III) Silent start point tu0 When a point in time when the above energy PwH suddenly changes is detected in the vicinity of the temporary start point tu', that is, within the time position from (tu/-5m5) to (tu'+20 ms). , that point is defined as the starting point tu. If it is not detected, the temporary starting point tu' Y starting point tu is used.

Note that when tv'>tu' and (tu+20 ms)>tv', the above range is from t (tu-5 ms) to tv.

IJV) Voiced starting point tv0 From the time position of tv' in the vicinity of the above tentative starting point tv' (
Within the time position range of tv+5m5), the above energy P
When a point in time when wL suddenly changes is detected, that point is defined as the starting point tv. If not detected, the temporary starting point t
Let v' be the starting point tv.

As described above, the respective starting points to and tv are extracted, but the above-mentioned point of sudden change in energy can be considered to be extracted as follows. That is, in the case of starting point tu extraction, 2 with a time width of 3 ms for the above energy PwH
For example, two observation windows are connected and scanned to take sample points at intervals of 0.5 ms.

Then, calculate the energy Pw at the first observation window and the energy PW2 at the second observation window, and when the value exceeds the threshold value 3.0, for example, at the time position corresponding to the connected position) , so that it can be assumed that something has changed. In addition, in the case of starting tv oil extraction,
Two observation windows having a time width of 3 ms are connected for the above energy PwL, and scanning is performed so as to take sample points at intervals of 0.5 ms. When the value of D P W corresponding to the above equation (4) exceeds the threshold value of 2.0, it is assumed that a sudden change has occurred at the time position corresponding to the connected position.

In this way, the voiced beginning tvo and the unvoiced beginning itu.

can be determined accurately. If the re-verification candidate is, for example, /ni/, then from table 10 there is no? The selection of the starting end is instructed, and the original voice energy waveform shown in FIG. 2(a) is cut out from the unvoiced starting point T2 shown in FIG. In the case of /g/, the original voice energy waveform shown in FIG. In this case, since the DP verification χ is performed by excluding some of the buzzers whose lengths fluctuate widely, the verification results are very good.

By the way, as mentioned above, for example, in /ni/←/g/, when both perform DP matching from the silent beginning, /ni/ will be compared with the longer part 'YDP by T2-T4, and these requests will be There is a disadvantage when choosing the one with the smaller distance in comparison. Therefore, among the respective starting points, use the slow ten thousand, and for /ni/, the voiced starting point T3 is used.
For v and /g/, if the silent starting point T4 is used to perform re-verification in the DP matching unit 11, the above-mentioned disadvantages can be greatly improved and recognition results with even higher accuracy can be obtained. Become.

〔Effect of the invention〕

In the present invention, /to/-/g/, /d/-
/Z/, etc., can these voiced and unvoiced beginnings be accurately detected and re-verified based on this? Therefore, the conventional D
Unlike P matching, which starts DP matching when a certain amount of power is obtained in the speech energy waveform, matching can be performed by accurately capturing the starting point of the consonant syllable.

Therefore, DP matching can be performed from the syllable-specific part, and as a result, highly accurate recognition results can be obtained.

[Brief explanation of the drawing]

FIG. 1 is an explanatory diagram of a voice recognition system, FIG. 2 is a voice power waveform diagram, and FIG. 3 is a configuration diagram of an embodiment of the present invention. In the figure, 1 is a low-pass filter, 2 is a mid-high-pass filter, and 3.4
1 is a power calculation section, 5 and 6 are threshold value determination sections, 7.8 is a starting edge detection section, 9 is a selector, 10 is a table, and 11 is a DPP combination section. Patent applicant Fujitsu Ltd. Representative Patent Attorney Haru Yamatani Sakae 12th layer

Claims (1)

[Claims] 1. Select a plurality of re-verification candidates obtained by collating previously registered monosyllabic speech and unknown input monosyllabic speech, and perform re-verification using an identification matching method suitable for the re-verification candidates. The speech recognition device to be compared is equipped with a medium-high-pass filter and a low-pass filter 7. Each filter output has a threshold value that changes depending on the other filter output and the point in time when the temporary threshold value Bi is exceeded. A start detection means is provided, which determines the point in time when the output exceeds each threshold as the start of a voiced consonant part and a start of an unvoiced consonant part, respectively, and a D'P matching means is provided. A monosyllable speech re-matching method characterized in that the starting point of the DP matching is set as the starting point of the voiced consonant part and the end of the unvoiced consonant part starting point. Monosyllabic speech re-verification method described in Section 1.