JPH0430598B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Technical Field of the Invention] The present invention relates to a speech recognition device that creates an optimal speech dictionary for each input speaker for speaker adaptation in a speech recognition system.

[Technology background]

In general, in speech recognition systems, recognition of a large number of words, for example, several hundred words or more, without any prior learning is still problematic in terms of recognition rate. On the other hand, a method in which a large number of words are all uttered in advance and registered is excellent in recognition accuracy, but it imposes too much burden on the speaker and is not practical. Therefore, a method is needed to create a dictionary suitable for the speaker by using as little pre-utterance data as possible.

One possible method is to select only one dictionary that is 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. There is a problem in that a large number of dictionaries must be prepared in order to ensure that there are dictionaries suitable for each speaker.

[Objects and Structure of the Invention] An object of the present invention is to reduce the burden on speakers in advance learning and to develop an optimal speech dictionary in a speaker adaptive method that optimizes a speech recognition system for each speaker and increases recognition accuracy. The purpose of this invention is to provide a means by which the system can be easily configured.

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, it will not work very well depending on the input speaker. There are cases where a suitable dictionary does not exist and therefore a high recognition rate cannot be obtained. In such a case, by using information from multiple dictionaries, it is possible to make the recognition rate higher than when any one of the dictionaries is used alone.

The present invention has been made by focusing on this point, and selects the main speech dictionary based on the recognition results using a small number of sample speech data as dictionaries, and selects a plurality of speech dictionaries with relatively high similarity. The above purpose is achieved by narrowing down to the dictionary, or by creating a new dictionary from it.

The present invention has the following features: (1) a plurality of speech dictionaries created by different speakers; a sample speech data group consisting of the same category group created for each of the plurality of speech dictionaries; A means of calculating the recognition rate by recognizing the voice data of the input speaker to be recognized using each data group as a dictionary, and selecting only the multiple voice dictionaries corresponding to the top multiple sample voice data groups that showed high recognition rates. The plurality of selected speech dictionaries are used as speech dictionaries for the input speaker.

(2) Recognizing a plurality of speech dictionaries created by different speakers, a sample speech data group consisting of the same category group created for each of the plurality of speech dictionaries, and each of the sample speech data groups as a dictionary. means for recognizing voice data of a target input speaker and calculating a recognition rate; means for selecting only a plurality of voice dictionaries corresponding to a plurality of top sample voice data groups showing a high recognition rate; means for creating a new voice dictionary by averaging a plurality of voice dictionaries, and using the voice dictionary created by averaging the plurality of selected voice dictionaries as a voice dictionary for the input speaker. This is a characteristic feature.

[Embodiments of the invention]

The details of the present invention will be explained below with reference to Examples.

In general, when using someone else's dictionary, the recognition rate is much lower than when using your own dictionary, but it can be improved through learning.

If you replace part of someone else's dictionary with your own dictionary to create a new dictionary, if the features used include a lot of information that represents individual differences, the recognition rate will increase depending on the amount of replacement. For example, a change like the solid line graph 1 in FIG. 1 is shown.

That is, when a small amount of someone else's dictionary is replaced with one's own dictionary, the recognition rate tends to decrease. This phenomenon occurs because when the feature amount includes a lot of personal information, another word uttered by the user has higher similarity than the same word uttered by another person. In this case, since it is possible to distinguish between one's own data and other people's data in the dictionary, a threshold for dissimilarity is set for one's own data, and one's own dictionary data exceeding this threshold is not adopted at the time of recognition. As a result, as shown in broken line graph 2 in FIG. 1, it can be seen that the recognition rate increases roughly in proportion to the amount of replacement.

In addition, dictionaries for many speakers are prepared in advance (two sets of majority vocabulary and minority vocabulary, i.e., sample vocabulary; the minority vocabulary may be part of the majority vocabulary), and the user can utter the sample vocabulary. hand,
Recognition is performed within the range of that vocabulary. the result,
If we decide to use a dictionary with a large number of vocabulary words written by the speakers of the sample vocabulary with the highest recognition rate, the average The recognition rate can be increased.

Instead of using one dictionary with the highest recognition rate, a method is used in which multiple dictionaries with high recognition rates are used.

Conventionally, the multi-template method and the average pattern method are often used as methods for using multiple dictionaries.

The multi-template method simply arranges a plurality of dictionaries in a two-dimensional manner and combines them into one dictionary. There will be multiple data (number of speakers) in one category,
At the time of recognition, processing is performed to find the most similar data among all of the data.

Next, in the average pattern method, feature values are averaged for each corresponding feature in a plurality of pieces of data in the same category, and a single piece of new data is created. In the case of audio, since there is a variation in time length, it is generally averaged after being correlated in the time direction.

In this embodiment, since the time length is normalized in the time direction by dividing the word length into 16 equal parts, the averaging operation can be easily performed.

FIG. 2 compares the effects of the multi-template method and the average pattern method.

Figure 2 shows the cases where the top 1, 3, 5, and 10 dictionaries with high similarity are selected from 40 dictionaries (200 words) for each of the 40 test speakers, and the vertical axis shows the top 1, 3, 5, and 10 dictionaries with high similarity. The axis shows the average recognition rate for 40 speakers. Graph 3 shows a case of a multi-template, and graph 4 shows a case of an average pattern dictionary. As is clear from the figure, the average pattern method is superior to the multi-template method. This is because, in the case of the average pattern method, personal information included in each dictionary is diluted by averaging, while effective feature information is 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 diagram is
The top 20 dictionaries with the highest similarity are selected from the 100 dictionaries (200 words) and arranged on the horizontal axis in order of similarity, and the five input speakers A, B, C,
The vertical axis is the recognition rate of D and E. For each input speaker, 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 represents the recognition rate for the top 3, 5, Represents the recognition rate averaged over 10 dictionaries.

In order to perform recognition without prior utterance, it is necessary to prepare a speaker-independent dictionary. As described above, it is not possible to obtain a high recognition rate if a dictionary of one specific person is used as a dictionary for unspecified speakers.

Here, we will consider the case where a plurality of given dictionaries are averaged for use by unspecified speakers.

It is assumed that dictionaries for 20 people (1000 vocabulary words) are registered in advance.

Also, regarding 20 input speakers different from the above,
Figure 4 shows the recognition rate data when recognition was performed using the following five types of dictionaries for 1000 vocabulary words.

A dictionary for non-specific speakers that is the average of 20 prepared dictionaries.Average of the main dictionary (1000 vocabulary) corresponding to the 10 sample dictionaries with high recognition rates after recognition was performed using a sample dictionary of 50 words. Dictionary Recognition was performed using a sample dictionary of 100 words, and the average of the main dictionaries (1000 vocabulary) corresponding to the 10 sample dictionaries with the highest recognition rate.The 10 highest recognition rates were recognized using a sample dictionary of 200 words. The average dictionary (1000 vocabulary) corresponding to the sample dictionaries The personal dictionary of 1000 vocabulary registered by input speakers' own utterances The graph in Figure 4 is the average of 20 input speakers. It represents the recognition rate.

By learning with sample dictionaries with a small number of vocabulary such as , , selecting multiple main dictionaries, and then creating an average dictionary, an individual improves in proportion to the number of pre-uttered words indicated by the dashed line 5 connecting and. The effect shown in characteristic 6, which is even higher than the recognition rate of the dictionary, can be achieved.

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 as an example the case of the speech dictionary configuration method based on the above-mentioned method.

FIG. 6 is a configuration diagram of the same, where 7 is an input section, 8 is a recognition section, 9 is a sample dictionary group file, 10 is a recognition result holding section, 11 is a selection section, 12 is a main dictionary group file, and 13 is an average dictionary. 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: a dictionary configuration mode and a recognition processing mode.

First, the mode changeover switch 15 is set to the sample dictionary group file 9 side to set the dictionary configuration mode. Here users can select a small number of words (100
utter words). Audio data based on this learning utterance is input from the input section 7 and recognized by the recognition section 8. The dictionary used at this time is one in the sample dictionary group file 9.

Sample dictionary group file 9 contains multiple (20 people)
Sample dictionaries uttered by speakers of (20)
, and all sample dictionaries consist of the same category group (100 vocabulary), and this category group includes all of the above-mentioned minority words for learning. This category group may be the same as or a part of the main dictionary group file 12, which will be described later, or may be unrelated.

The recognition result holding unit 10 holds recognition results for each sample dictionary in the sample dictionary group file 9.

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 12 stores a dictionary in which recognition target word groups (1000 vocabulary) are registered as utterances by a plurality of speakers who have uttered the sample dictionary data of the sample dictionary group file 9.

The average dictionary creation unit 13 extracts main dictionaries written by the same speaker as the plurality of sample dictionaries selected by the selection unit 11 from the main dictionary group file 12, and creates one dictionary by averaging these dictionaries. The created average dictionary is stored in the average dictionary storage section 14.

Here, the setting of the mode changeover switch 15 is changed to the average dictionary storage section 14 side, and the recognition processing mode is set. At the time of use, the speech input from the input 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.

Note that when a multi-template method is used instead of the average pattern method, a plurality of dictionaries selected from the main dictionary group may be directly used as dictionaries for recognition processing.

〔Effect of the invention〕

As described above, according to the present invention, from a plurality of dictionaries prepared in advance, it is possible to create a dictionary with a higher recognition rate than any of them depending on the speaker. You can quickly and easily create high-precision dictionaries without having to

[Brief explanation of drawings]

Figure 1 is an illustration of the learning effect for other people's dictionaries, Figure 2 is an illustration of the effect of the number of selected dictionaries for multi-template dictionaries and average pattern dictionaries, and Figure 3 is an illustration of the effect of the similarity ranking of selected dictionaries. An explanatory diagram, FIG. 4 is an explanatory diagram of the learning effect for a dictionary for unspecified speakers, FIG. 5 is an explanatory diagram summarizing FIGS. 1 to 4, and FIG. 6 is an example system of the present invention. FIG. In the figure, 7 is an input section, 8 is a recognition section, 9 is a sample dictionary group file, 10 is a recognition result holding section, 11 is a selection section, 12 is a main dictionary group file, 13 is an average dictionary creation section, and 14 is an average dictionary A storage section, 15 a mode changeover switch, and 16 an output section.

Claims (1)

[Claims] 1. A plurality of speech dictionaries created by different speakers, a sample speech data group consisting of the same category group created for each of the plurality of speech dictionaries, and each of the sample speech data groups. means for recognizing speech data of an input speaker to be recognized using a dictionary and calculating a recognition rate; and means for selecting only a plurality of speech dictionaries corresponding to a plurality of top sample speech data groups showing a high recognition rate. A speech recognition device characterized in that the plurality of selected speech dictionaries are used as speech dictionaries for the input speaker. 2 A plurality of speech dictionaries created by different speakers, a sample speech data group consisting of the same category group created for each of the plurality of speech dictionaries, and a recognition target using each of the sample speech data groups as a dictionary. means for recognizing voice data of an input speaker and calculating a recognition rate; means for selecting only a plurality of voice dictionaries corresponding to a plurality of top sample voice data groups showing a high recognition rate; and means for creating a new speech dictionary by averaging the speech dictionaries, and the speech dictionary created by averaging the plurality of selected speech dictionaries is used as the speech dictionary for the input speaker. voice recognition device.