JPS6076800A - Voice recognition 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 The present invention relates to a speech recognition method using pattern matching technology.

As a new pattern matching method for recognizing word sounds, the applicant first developed two types of two-dimensional patterns (local peak trajectories) based on a spectrum graph with frequency and time as variables. Thinning pattern and 2
We have proposed various methods for calculating similarity using the degree of overlap of digitized broad patterns).

To explain in more detail, for the purpose of simple voice recognition,
A single audio signal taken in from a microphone is led to a feature extraction unit, and frequency analysis is performed at time intervals of several msec to several tens of m5ec by some method, and the short-term power spectrum of the audio signal is obtained, and this is extracted as a feature. When expressed as a quantity, a pattern in which characteristic quantities such as stiffness are displayed on a two-dimensional plane with frequency and time as two axes is known as a time spectrum pattern of audio. The present invention is based on this time spectrum pattern, and has two types: a thin line pattern (Figure 1) in which the local peaks are connected, and a broad pattern (Figure 2) that is binarized by setting a certain threshold value. This method involves calculating the degree of similarity by superimposing and comparing the patterns of
When performing matching processing on different types of patterns, it is necessary to align each time axis and frequency for comparison. However, even for the same word by the same speaker, the length of the dictionary pattern to be matched and the time axis of the input pattern differ due to variations in the rate of speech for each utterance, and matching cannot be performed as is. Therefore, conventionally, a linear expansion/contraction matching method has been used in which the time axis of either the dictionary or the input time spectrum pattern is linearly expanded/contracted to match the length of the time axis of the other pattern to calculate the similarity. However, the variation in the speaking speed of word sounds is not necessarily linear, and the speed variation differs depending on the sounds that make up the word, so linear matching has not been able to obtain a sufficient recognition rate.

Moreover, the present invention was made in view of the above-mentioned circumstances,
In particular, for each frame of the dictionary pattern and input pattern, it is determined whether the frame is voiced (V) or unvoiced (UV), and the dictionary patterns are clustered based on the time series of V and UV to reduce calculation time. The aim is to improve the recognition rate.

In addition, the structure of the present invention will be described below based on examples.

In the present invention, when registering word dictionary patterns, it is determined whether each frame is V (voiced sound) or UV (unvoiced sound) for each pattern, and V or UV is recorded for each frame in advance. However, various methods can be considered for this determination.For example, as for waveform processing, division determination is made by utilizing the fact that the periodic structure corresponds to V and the random part corresponds to UV. In addition, processing in the frequency domain takes advantage of the property that the slope of the approximation straight line of the spectral envelope is negative for V (voiced sounds) and O or positive for UV (unvoiced sounds). It can also be classified.

Figure 3 is a diagram showing an example of marking each dictionary pattern with ■ or UV as described above. In Figure 1, (a) is the frame number, (b)
(c) shows the labeling, and (c) shows the change pattern. In this way, the changes in V and UV are recorded in advance together with the dictionary pattern as one of the features of this word. During recognition, V and UV classification is similarly performed for each frame of the input pattern, and labeling is applied to create a V and UV change pattern. Next, a preliminary selection is performed to select a pattern that changes similarly to this change pattern from among the change patterns in the dictionary, and by performing such preliminary selection, the similarity calculation can be shortened and the recognition rate can be improved. be able to. Furthermore, as a final step, a matching calculation is performed between a plurality of pre-selected dictionary patterns and the input pattern, and a recognition result is output.

[Example 1] When matching an input cover turn and a plurality of preselected dictionary patterns, either the input cover turn or the dictionary pattern is linearly expanded or contracted to match the length of the time axis with the other pattern, Perform similarity calculation.

[Example 2] Linear expansion and contraction is performed during pattern matching between the input cover pattern and the preselected pattern, but in calculating the similarity, the correspondence between the input cover pattern and the dictionary pattern is determined as shown in Fig. 4. The similarity of each frame is calculated only when the labels of the respective frames are both V (voiced sound) or both UV (unvoiced sound). Similarity calculation is not performed for frames in which opposing frames have different labels, such as one label being V and the other UV. By taking such measures, it is possible to avoid calculating the degree of similarity between different types of phonemes due to speed fluctuations during utterance.

[Embodiment 3] During pattern matching of the input cover pattern and the preselected dictionary pattern, as shown in FIG.

According to each change pattern, partial linear expansion/contraction is performed in which the lengths are expanded/contracted between parts of the same label to calculate the degree of similarity. In this case, non-linear expansion and contraction is performed when looking at the entire word, and of course matching calculations between elements of different types can be avoided.

Example - As is clear from the above description, according to the present invention,
By creating change patterns for voiced sounds) and UV (unvoiced sounds) and using dictionary patterns based on these V and UV change patterns, it is possible to shorten the similarity calculation time and improve the recognition rate.

[Brief explanation of drawings]

FIG. 1 shows a thinning pattern, FIG. 2 shows a broad pattern, and FIGS. 3 to 5 each show an embodiment of the present invention. ■...Voiced sounds, UV...Unvoiced sounds. → No. 31! I a (r-me No.) + 2 3 4 5 s 7・-
-----------Nb (Uni ring) v v v v
v uv v v uvuvC (liquefied putter>') v
uv v uvFigure 4Figure 5c (uH,;,<,y;?) v v v uv uv
uv uv v uv uv uv Procedural amendment (Revised 1. Indication of the case 1982 Patent Application No. 185671 2, Benefits of the invention/Speech recognition method 3, Relationship with the case to be amended Patent applicant Otaku Nakamagome address Tokyo 1-3 Nakamagome, Ota-ku, Tokyo @6 Name (Name) (674) Ricoh Co., Ltd. Representative Hama 1)
Hiro Chatelaine Inn Yokohama No. 807 No. 5, date of amendment order 7, contents of amendment (1), "single elimination" stated in page 2, line 13 and line 14 of the specification is amended to "word". (2) ``Correct the sound j to ``phoneme'' as described on page 3, line 19.

Claims (3)

[Claims] (1) A label of either a voiced sound or an unvoiced sound is attached to each frame of both the dictionary and the input kataan, and preliminary selection of the dictionary is performed based on the voiced/unvoiced change pattern based on this label. Voice recognition method. (2) A patent claim characterized in that when matching an input cover pattern and a dictionary pattern, the voiced/unvoiced label for each frame is used to avoid matching calculations between different phonemes. The speech recognition method described in scope (1). (3) Partial linear expansion/contraction matching in which when matching input cover patterns and dictionary patterns, matching is performed by linearly matching the lengths of corresponding parts of the same label using the voiced/unvoiced labels for each frame. A speech recognition method according to claim (1), characterized in that the method performs the following.