JPS602998A - voice recognition 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] [Technology of the Invention] The present invention relates to a speaker adaptation method in a speech recognition system, and in particular to a method for selecting an optimal speech dictionary for each specific speaker from among a plurality of speech dictionaries. 0 Regarding Speech Dictionary Construction Method [Technical Background] In general, in speech recognition systems, it is difficult to recognize a large number of filler words, for example, several hundred words or more, without any prior learning. There is a problem with this. On the other hand, a method in which a large number of words are all uttered in advance and registered is excellent in terms of recognition accuracy, but it imposes too much burden on the speaker and is not practical. Therefore, there is a need for a method of creating a dictionary suitable for the speaker by using as little pre-utterance data as possible.

One possible method is to select only one dictionary most suitable for the speaker from among multiple dictionaries prepared in advance, for example, dictionaries each created based on the sounds uttered by multiple other speakers. However, in order to guarantee the existence of a dictionary suitable for each speaker, there is a problem in that a large number of dictionaries must be prepared.

[Object and structure of the invention]

An object of the present invention is to reduce the burden on speakers in advance learning and easily construct an optimal speech dictionary in a speaker adaptive method that optimizes a speech recognition system for each speaker to improve recognition accuracy. The goal is to provide the means to do so.

In general, even if you select and use only one dictionary that is optimal for an input speaker from among a plurality of given speech dictionaries, if the number of original speech dictionaries is small, depending on the input speaker, At
b) A suitable dictionary may not be transferred, and therefore a high recognition rate may not be obtained. In such a case, by using information from multiple dictionaries, it is possible to increase the recognition rate higher than when using any of the dictionaries alone. The present invention was made with attention to this point. Alternatively, the main speech dictionary can be selected based on the recognition results using a small number of sample speech data as a dictionary, and the speech dictionaries with relatively high similarity can be narrowed down to a plurality of speech dictionaries, or even further. The above objective is achieved by creating a new dictionary.

The present invention has the following features: (1) In a speech recognition system equipped with a plurality of speech dictionaries, a sandal speech data group consisting of the same category group created for each of the plurality of speech dictionaries, and the sample speech A method for calculating the degree of similarity between a data group and the speech data of a specific speaker to be recognized, and a method for selecting only a plurality of speech dictionaries corresponding to a plurality of top group speech data groups showing a high degree of similarity. The selected plurality of speech dictionaries are used as speech dictionaries for the specific speaker.

12) A voice recognition system equipped with multiple voice dictionaries,
means for calculating a degree of similarity between a sample voice data group made up of the same category group created for each of the plurality of voice dictionaries, the sandals voice data group and voice data of a specific speaker; Means for selecting only a plurality of speech dictionaries to correspond to a plurality of top sample speech data groups showing high similarity, and means for creating a new speech dictionary by averaging the plurality of selected speech dictionaries. The present invention is characterized in that a speech dictionary created by averaging the plurality of selected speech dictionaries is used as a speech dictionary for the specific speaker.

[Embodiments of the invention]

The details of the present invention will be explained below based on examples.

In general, if you use someone else's dictionary, your Iga knowledge rate will be much lower than if you use your own dictionary, but you can improve it through learning. If a new dictionary is created by replacing ``,'' and the features used include a lot of information that represents individual differences, the recognition rate will change depending on the amount of replacement, for example, as shown in the solid line graph l in Figure 1. It shows a change like this.

In other words, if you replace a small amount of someone else's dictionary with your own dictionary, the recognition rate tends to decrease.
This phenomenon occurs because when the features contain a lot of personal information, the similarity between another foreign word uttered by oneself becomes higher than the same foreign word uttered by another person. It is. In this case, ■ Since it is possible to distinguish between one's own data and other people's data in the dictionary, a threshold regarding dissimilarity is set for one's own data, and the own dictionary that exceeds this threshold at the time of recognition. By not adopting the data, it is possible to see an increase in the recognition rate roughly proportional to the amount of replacement, as shown in broken line graph 2 in FIG.

■ In addition, dictionaries for multiple speakers are prepared in advance (two sets of multiple target and minority target, i.e., sample target; the minority target can be a part of the multiple target), and the user uses the sample words mentioned above. Speak out the proposal and recognize it within the scope of the counter proposal. As a result, if we decide to use a dictionary with many countersuggestions by speakers of the sample word Kume, which had the highest recognition rate, we can simply
The average recognition rate can be increased compared to the case where a dictionary of people is prepared and all users use that dictionary.In addition to using one dictionary with the highest recognition rate in 0 ■ ■, υ There is a method of using multiple words tt with a high recognition rate.

Conventional methods of using multiple dictionaries include the multiten rate method, the average no J? The multi-template method, in which the tongue method is often used, simply arranges multiple dictionaries in a plane and combines them into a single dictionary. Since there are multiple pieces of data (number of speakers) in one category, at the time of recognition, a process is performed to find the most similar data among all the data. is the average of the feature values for each corresponding feature in multiple data in the same category to create one new data.In the case of 0 voice, i,
Since the time length varies, it is generally averaged after mapping in the time direction. In this embodiment, the time length is normalized by dividing the word length into 16 equal parts. Therefore, the averaging operation can be easily performed.

FIG. 2 compares the effects of the multiten rate method and the average/9 tan method.

Figure 2 shows the top 1 most similar words from 40 dictionaries (200 words) for each of the 40 test speakers.
, 3, 5, and 10 dictionaries are selected on the horizontal axis, and the vertical axis shows the average recognition rate of 40 speakers. If graph 3 is a multi-tenure rate, graph 4
It is clear from the figure that the average pattern method is superior to the multi-template method as shown in the figure where This is because each personal information portion is diluted by averaging, and on the other hand, effective characteristic information portions are emphasized. On the other hand, in the case of the multi-template method, such an effect cannot be produced.

FIG. 3 shows typical example data to further clarify the effect of the average pattern method. The figure shows 100 dictionaries (2
The top 20 dictionaries with the highest similarity from 00 words) are arranged on the horizontal axis in order of similarity, and the recognition rates of the five input speakers A, B, C, D, and E are plotted vertically. For each large input user, the solid line graph extending downward represents the recognition rate when each of the top 20 dictionaries is treated as a single dictionary, and the dotted line graph extending upward The graph represents the recognition rate averaged over the top 3, 5, and 10 dictionaries.

■ In order to perform recognition without prior utterance, it is necessary to prepare a speaker-independent dictionary. As mentioned above, it is not possible to obtain a high recognition rate if one specific person's dictionary is used as a dictionary for non-specific speakers.Here, when a plurality of given dictionaries are averaged and used for non-specific speakers. 0 It is assumed that 20 people's dictionaries (1000 words!) have been registered in advance.0 Also, regarding the addition of input speakers other than the above, the following 5 types of dictionaries are used to recognize 1000 words. Figure 4 shows the recognition rate data when performing recognition using a 50-word Sanzor dictionary. The main dictionary corresponding to the sample dictionary (1000
Word translation) Kaneko's average dictionary ■ Recognition is performed with a sample dictionary of 100 words, and the main words corresponding to the 10 sample dictionaries with high recognition rates - 1: (1000 word ideas) Kaneko's average dictionary @ 200 The main dictionary (10
00 word search) Kaneko's Average Dictionary■ A personal dictionary of 1000 words recorded by input speakers' own utterances.The graph in FIG. 4 shows the average recognition rate for 20 input speakers.

The pre-utterance shown by the dashed line 5 connecting ■, ■, and ■ is to study with a sample dictionary of minority word ideas, select multiple main dictionaries, and then create an average dictionary. It is possible to achieve the effect shown in characteristic 6, which further exceeds the recognition rate of personal dictionaries, which improves in proportion to the number of words.

FIG. 5 is a graph summarizing the above-mentioned relationships.

Next, the configuration of one embodiment of the speech recognition system according to the present invention will be described using the above-mentioned @-based speech dictionary construction method as an example.

FIG. 6 is its configuration diagram, where 7 is an input section, 8 is a recognition section,
9 is a sample dictionary group file, 10 is a recognition result holding unit,
11 is a selection section, 12 is a main dictionary group file, 13 is an average dictionary creation section, 14 is an average dictionary storage section, 15 is a mode changeover switch, and 16 is an output section.

The system of this embodiment operates in two modes: dictionary configuration mode and recognition processing mode. First, set the mode changeover switch 15 to the sample dictionary group file 9 side to switch to dictionary configuration mode. utters a small number of learning words (100 words) 0 Voice data based on this learning utterance is input manually from the input unit 7 and is recognized by the recognition unit 8 0 The dictionary used at this time is Sandals. O in the dictionary group file 9 The sample dictionary group file 9 includes sample words -t (20 pieces) uttered by a plurality of (20) speakers,
And all sample dictionaries have the same category group (100
This category group is the same as or a part of the main dictionary group file 12, which will be described later. The recognition result holding unit 10 stores the sample dictionary group file 9.
Recognition results are retained for each sample dictionary in the dictionary.

The selection unit 11 selects a sample dictionary with a high recognition rate from among the sample dictionaries in the sample dictionary group file 9.

The main dictionary group file n contains the sample dictionary group file 9.
The sample dictionary data of 7'c is uttered and recognized by multiple speakers 11. A dictionary in which word groups (1ooo words) are registered as utterances is stored.

The average dictionary creation unit 13 extracts the main dictionaries by the same speaker as the plurality of sample dictionaries selected by the selection unit 11 from the main dictionary group file 12, and averages these dictionaries.
The created average dictionary is stored in the average dictionary storage unit 14. Here, the mode changeover switch 1
5 is set to the average dictionary storage unit 14 side, and the recognition processing mode is set. At the time of Fang U, the voice input from the human power section 7 is recognized by the recognition section 8 using the dictionary stored in the average dictionary storage section 14, and the result is output from the output section 16. If the multi-template method is used instead of the average slam method, the plurality of dictionaries selected from the main dictionary group may be used as dictionaries for recognition processing. [Effects of the Invention] As described above, the present invention According to Arara25≧Multiple dictionaries, it is possible to create a dictionary with a higher recognition rate than any of the prepared dictionaries depending on the speaker, so it is possible to quickly and easily create a dictionary that has a higher recognition rate than any of them, depending on the speaker. You can easily create high-precision dictionaries.

[Brief explanation of the drawing]

Figure 1 is an explanatory diagram of the learning effect on other people's dictionaries, and Figure 2 is an illustration of the learning effect on other people's dictionaries.
Figure 3 is an illustration of the effect of the similarity ranking of selected dictionaries, Figure 4 is an illustration of the learning effect for dictionaries for non-specific speakers, and Figure 5 is an illustration of the effect of the number of selected dictionaries. is an explanatory diagram summarizing FIGS. 1 to 4, and FIG. 6 is a configuration diagram of a system according to an embodiment of the present invention. In the figure, 7 is an input section, 8 is a recognition section, and 9 is a sandal dictionary group. 10 is a recognition result storage unit, 11 is a selection unit, ν is a main dictionary group file, 13 is an average dictionary creation unit, 14 is an average dictionary storage unit, 15 is a mode changeover switch, and 16 is an output unit.Patent application Person Fujitsu Ltd. Representative Patent Attorney For long diameter Hiroshi Fumi (1 other person) 1%) Approved.ーーー“-・I1c←−Ichigohan Hajime

Claims (2)

[Claims] (1) In a speech recognition system equipped with multiple speech dictionaries, a sample speech data group consisting of the same category group created for each of the plurality of speech dictionaries, and identification of the sample speech data group and the recognition target. The method includes a means for calculating the degree of similarity between the voice data of the speaker and a means for selecting only a plurality of voice dictionaries corresponding to a plurality of top sample voice data groups showing a high degree of similarity. A speech dictionary configuration method characterized in that a plurality of speech dictionaries are used as speech dictionaries for the specific speaker. (2) In a speech recognition system equipped with a plurality of speech dictionaries, a sandal speech data group consisting of the same Qadiri group created for each of the plurality of speech dictionaries, and the sandal speech data group and the speech data of a specific speaker. means for calculating the degree of similarity between the plurality of voice dictionaries, means for selecting only the plurality of voice dictionaries corresponding to the top plurality of sandal voice data groups showing a high degree of similarity, and means for averaging the plurality of selected voice dictionaries. and a means for creating a new speech dictionary using the plurality of selected speech dictionaries, and the speech dictionary created by averaging the plurality of selected speech dictionaries is used as the speech dictionary for the specific speaker. 0