JPH0293696A - Speech recognition device - Google Patents

Abstract

PURPOSE:To perform accurate speech recognition by setting a free area which is set before a temporary start point in terms of time longer than a free area which is set after the temporary start point when the start point and end point of a partial pattern are set. CONSTITUTION:This device is equipped with a microphone 1, a speech analysis part 2, an input speech pattern buffer 3, a speech section segmentation part 4, an asymmetrical end point free DP matching part 5, and a standard speech pattern memory 6. In this case, the free area set before the head of a part segmented as a speech section candidate in terms of time is set longer than the free area set after said head and the free area set after the tail of the part segmented as the speech section candidate in terms of time is set longer than the free area set before said tail. Consequently, a language section in noises and continuous speech is accurately segmented and the recognition rate can be improved.

Description

DETAILED DESCRIPTION OF THE INVENTION (A) Field of Industrial Application The present invention relates to a speech recognition device that performs accurate speech recognition by accurately cutting out speech sections from input speech.

(b) Conventional technology In speech recognition, a practical speech recognition device that cuts out speech sections from input speech, such as recognition in noise or phoneme recognition in continuous speech, is based on the power of human speech. but,
A section that exceeds a certain threshold value is cut out as a speech section candidate, a temporary start and end are determined, and the cut out partial pattern and the standard speech pattern are used as a nonlinear method with the temporary start and end of the input speech pattern side free. In many cases, the starting end and ending end of the partial pattern are determined by comparing by matching.

FIG. 2 shows an example of such a conventional speech recognition device.

The voice input from the microphone [7] is analyzed by the voice analysis unit [8] and converted into a spectrum or cepstrum parameter time series at a frame period of about 10 ms. This parameter time series is stored in the input voice pattern buffer [9]. The speech section extraction unit [10] cuts out sections in which the power of the input speech is equal to or higher than a certain threshold value (TH) as speech section candidates, and performs endpoint DP matching between the speech section candidate information and the parameter time series. Department [Send to 11th.

The operation of this end point free DP munching unit [+1] is as follows.

That is, based on the data sent from the voice section extraction unit [10], a free area FBb taken before the temporary starting point in the time direction and a temporary starting point as shown in FIG. 3(a) are determined. The free area FBa, which is taken later in the time direction, is set to the same time length. Furthermore, the free area FAb, which is located before the temporary end in the time direction, and the free area JAF, Aa, which is taken after the temporary end in the time direction, are set to the same time length.

End point free DP using the free area obtained in this way
By munching, mapping is performed between the standard voice pattern in the standard voice pattern memory [12] and the input voice pattern.

(c) Problems to be Solved by the Invention In the conventional speech recognition device described above, the free areas at the temporary start and end points are FBa=FBb, F, Aa
= F Ab, a problem as shown in FIG. 4 occurs.

That is, for example, it is assumed that one word "Aikagi" and "Squid" is stored in 12 standard pattern voice memories.

I just input the word "Aikagi" through the microphone [7], but as shown in Figure 4 (a), the power at the beginning and end of the word is small, so the voice candidates are determined by the power at the threshold (TH). The result of cutting out the section is shown in the same figure (b).
become that way.

When matching is performed by applying the free area shown in Figure 3(a) to this voice candidate section, the input voice word ``Aikagi'' is matched with the word ``Ika'' with the beginning and end of the word deleted. It comes off. As a result, the matching distance with "squid" in the same figure (c) is higher than that of the "squid" in the same figure (c).
It will be smaller than that of the "Ai-key" in d), leading to erroneous recognition.

(d) Means for Solving the Problems The speech recognition device of the present invention cuts out a section in which the voice power exceeds a certain threshold value as a speech section candidate, and compares the cut out partial pattern with the standard speech pattern. When determining the start and end of the partial pattern by comparing by non-linear matching with the temporary start and end of the input audio pattern free, as shown in Figure 3 (b), from the temporary start Set the free area FBb to be taken before the temporary start point to be longer than the free area FBa to be taken after the temporary start point in the time direction, and set the free area FAa to be taken after the temporary end point in the time direction to be earlier than the temporary end point in the time direction. This is set to be longer than the free area FAb.

(e) Effect In the speech recognition device of the present invention, by setting the free area as FBb>FBa and FXa>FAb as shown in FIG. 4(e), ) Even under the same input voice conditions as shown in Figure 3(f), the matching distance with ``squid'' in Figure 2(f) can be increased by an amount corresponding to the area shown by the hatching in Figure 3(eJ). .Therefore, the 7inch distance from the "Ai-key" shown in FIG.

(f) Embodiment FIG. 1 shows an embodiment of the speech recognition device of the present invention. The voice input from the microphone [1] is analyzed by the voice analysis unit [2] and converted into a spectrum or cepstrum parameter time series at a frame period of about 10 m5. This parameter time series is the input speech pattern buffer [
3]. The speech section extraction unit [4] cuts out sections in which the power of the input speech is equal to or higher than a certain threshold value (TH) as speech section candidates, and uses the speech section candidate information and the parameter time series as an asymmetric endpoint filter J-DP. Send it to the matching section [5].

The asymmetrical endpoint IJ-DP matching unit [5], which is the most characteristic feature of the device of the present invention, performs the following process based on the data sent from the voice section extraction unit [4], as shown in Fig. 3(b). A free area FB taken in front of the temporary starting point in the time direction.
b, and the free area F taken after the temporary starting point in the time direction.
Set Ba so that F Bb > F Ba, and set the free area F Ab before the temporary end in the time direction.
and the free area FA taken after the temporary end in the time direction.
a and F so that Aa>FAb. Furthermore, end point free DP munching processing is performed in the free area set under these conditions to match the standard voice pattern in the standard voice pattern memory [6] with the input voice pattern.

(G) Effects of the Invention As is clear from the above description, the speech recognition device of the present invention prevents local pattern matching errors in endpoint-free DP matching and accurately cuts out word sections in noise or continuous speech. The recognition rate can be improved.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of the speech recognition device of the present invention, FIG. 2 is a block diagram of a conventional speech recognition device, and FIG. 3(a)
(b) and Figure 4) Ushi I Nakai is a voice pattern diagram. [1],,,? Iklofon, [2], Speech analysis section, [3], Input speech pattern buffer, [4],...
,Speech section extraction unit, [5] ,,Asymmetric end point free DP matching unit, [
6] ,,,Standard voice pattern memory.

Claims (1)

[Claims] (1) Among the standard speech pattern extracted in advance by the speech analysis means and the input speech pattern extracted by the speech analysis means, sections in which the speech power exceeds a certain threshold are selected as speech section candidates. In a speech recognition device, the cut out partial pattern and the standard speech pattern are compared by non-linear matching with the starting end and ending end of the input speech pattern side free, and determining the true starting end and ending end of the partial pattern, The free area taken before the beginning of the part cut out as a speech section candidate in the time direction is
A free area that is set longer than a free area taken after the beginning of the part cut out as a speech section candidate in the time direction and a free area taken after the end of the part cut out as a speech section candidate in the time direction is cut out as a speech section candidate. 1. A speech recognition device comprising non-target end point free matching means for setting a free area longer than a free area taken temporally before the end of a part of the target part.