JPS63289599A - Segmentation apparatus - 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 Industrial Application) The present invention relates to an improvement of a segmentation device for speech recognition of unspecified speakers.

(Prior art and its problems) In methods of speech recognition using segments smaller than words, such as syllables or phonemes, the accuracy of dividing input speech into segments, that is, the accuracy of segmentation, has a significant effect on the recognition rate. If the recognition target is a specific speaker, it is possible to use features that depend on each speaker, but if the target is an unspecified speaker, the acoustic features differ for each speaker. Conventional segmentation devices that divide input speech into segments based only on the acoustic characteristics of the input speech sometimes result in extreme segmentation errors. Conventional devices have had such problems.

Therefore, the present invention attempts to avoid extreme segmentation errors and perform segmentation with high accuracy by estimating in advance the range in which segment boundaries exist.

(Means for Solving the Problems) Means provided by the present invention to solve the above-mentioned problems is an apparatus that performs speech recognition using segments smaller than words, such as syllables or phonemes, means for extracting acoustic feature parameters from speech; means for storing feature parameters, segment boundaries, and features of the segment boundaries as word standard patterns; means for determining a point in time of the input speech that corresponds to the segment boundary of the standard word pattern in a temporal manner; and searching for a segment boundary of the input speech within a certain range from the point in time using characteristics of the segment boundary. It is characterized by comprising means.

(Operation) Input speech is converted into a time series of feature parameters by the acoustic feature parameter extraction means. Here, the feature parameters are parameters that reflect acoustic characteristics.0 Journal of the Acoustical Society of Japan Vo 127, No. 1, p.
1971-09), Sakoe, Chiba, "Continuous word recognition based on temporal normalization of speech using dynamic programming", the slam matching method (hereinafter referred to as DP matching)
By using , it is possible to temporally associate the feature parameter time series of input speech with pre-registered word standard patterns. Therefore, if you store the results of dividing the standard pattern into segments in advance, it is possible to store the results of dividing the standard pattern into segments.
(referred to as segment boundary candidates) can be found. Since the actual segment boundary is considered to exist in the vicinity of the segment boundary candidate obtained in this way, segmentation can be efficiently performed by searching for the segment boundary in a separately determined section that includes the segment boundary candidate. Furthermore, when searching for a segment boundary near a segment boundary candidate, since the segment to be searched for is already known, the segment boundary can be found by selecting appropriate feature parameters. In the case of specific speakers, it is of course possible to perform word recognition using the distance from the word standard pattern obtained as a result of DP matching, but in the case of unspecified speakers, sufficient recognition accuracy cannot be obtained. Therefore, only the temporal correspondence between the input voice and the standard pattern is performed here. Therefore, it is only necessary to store a smaller number of feature parameters than when performing word recognition, and the amount of storage and calculation can be reduced.

(Example) Fig. 1 is a block diagram showing an embodiment of the present invention, Fig. 2 is a diagram illustrating a method for temporally associating input speech and standard patterns, and Fig. 3 is a diagram illustrating a method for temporally associating input speech and standard patterns. FIG. 3 is a diagram for explaining an example of a determination method. The operation will be briefly explained below.

Now, the input voice “3” is divided into three segments /s/, /a
Segmentation is performed into /, /N/. The input speech is converted into a time series of feature parameters in the acoustic analysis section 1. The feature parameters here may be any parameters that reflect acoustic features, such as the output of a filter bank, mel cepstral coefficients, number of zero crossings, formant frequency, etc. The feature parameters of each word, segment boundaries, and features of the segment boundaries are registered in advance in the storage unit 2 as a word standard pattern.The matching unit 3 reads the word standard pattern from the storage unit 2 and matches it with the output of the analysis unit 1. I do.

After optimization is performed by DP matching, by tracing the bus that gives the optimal value, it is possible to associate the input voice with the standard pattern, and to find segment boundary candidates. Figure 2 shows the temporal correspondence.
In Figure 2, 21 represents the energy envelope of the word standard pattern, 22 represents the energy envelope of the input speech, 23 represents the matching plane, and 24 represents the path that gives the optimal value. represent.

There are various methods for determining the interval to search for segment boundary candidates. Assuming that the segment boundary candidate is at time i, 1) A method of setting it to a fixed value.

For a certain value j, let the interval be [i-j, i+j].

2) A method that reflects the difference in time length from the standard pattern in 1).

When the input voice has a duration α times that of the standard pattern, the interval is set to [1−jXα, i+jXα].

etc. are possible.

When searching for a segment boundary candidate in the interval [il, i2] defined in this way, it is possible to take into consideration the acoustic characteristics of the segment boundary candidate to be searched for. That is, in the previous example, when searching for the boundary between segments /s/ and /a/, 3 kH2 for segment /S/
Based on the knowledge that there is a concentration of energy in the above frequency range, we calculated the amount of change in the ratio of energy above 3 kHz to the energy in the entire frequency range, and found that it was the minimum value in the interval [12, i2]. It is only necessary to find the time point taken as a segment boundary, and the storage unit 2 stores the characteristics of such a segment boundary.

The boundary determining unit 4 determines a search interval from the result of the matching unit 3,
The characteristics of the segment boundaries are read from the storage unit 2, and the segment boundaries are determined from the analysis results from the analysis unit 1. Third
In the figure, 31 indicates the energy envelope of the input voice, and 3
2 is 3 kHz of energy in the entire frequency range
In FIG. 3, which shows the amount of change in the proportion of energy equal to or higher than z, i is a segment boundary candidate, and i' is an interval [il, i
2], the segment boundary is determined as the point in time when the amount of change in the ratio of energy of 3 kHz or more to the energy in all 32 frequency regions takes a minimum value.

(Effects of the Invention) As described above, according to the device according to the present invention, speech of unspecified speakers can be segmented with high accuracy.
It becomes possible to realize a highly accurate speech recognition device.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of the present invention, FIG. 2 is a diagram illustrating a method for temporally corresponding input speech and standard patterns, and FIG. 3 is an example of a method for determining segment boundaries according to the present invention. FIG. In the figure, 1 is an acoustic analysis unit, 2 is a word standard pattern storage unit, 3 is a matching unit, 4 is a boundary determination unit, 21 is an energy envelope of input speech, 22 is an energy envelope of a word standard pattern, 23 is a matching plane, 24 is matching pass,
31 is the energy envelope of the input voice, and 32 is the amount of change in the ratio of energy of 3 kllz or more to the energy in the entire frequency range.

Claims (1)

[Claims] In a device that performs speech recognition using segments smaller than words such as syllables or phonemes, a means for extracting acoustic feature parameters from input speech, and storing feature parameters, segment boundaries, and features of the segment boundaries as word standard patterns. means for temporally associating the input speech and the word standard pattern using the acoustic feature parameter to determine a time point in the input speech corresponding to the segment boundary of the word standard pattern; and means for searching for a segment boundary of the input speech within a certain range from the point in time using boundary characteristics.