JPS6370296A - Word registration - Google Patents

Abstract

(57) [Abstract] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed Description of the Invention] [Summary] When registering words of a specific speaker for word speech recognition, the speech feature parameters of the words uttered for registration are already stored in memory and the speech features of the registered words are stored. Compare it with the parameters, and if there are any words that are similar, have them say the word emphasizing the difference from similar words, or say a word that has the same meaning in a different way, and then register it again. By doing so, the recognition rate during word speech recognition is improved by not registering words that have a similar degree of similarity to the speech feature parameters of words that have already been stored and registered in the memory.

[Industrial application field]

The present invention relates to an improvement in a word registration method for registering words of a specific speaker for word speech recognition of a specific speaker.

Word speech recognition devices for specific speakers can be made compact using hour wheels, so they are used in many devices.

In order to use this J ■ language speech recognition device, it is required that there are few false recognitions.

Although efforts are currently being made to improve misrecognition, they have reached their limits to some extent, and there is a demand for a word registration method that can further reduce misrecognition.

[Conventional technology]

A conventional example will be explained below using figures.

FIG. 3 is a professional diagram of a conventional word registration device, and FIG. 4 is a block diagram of an example of a speech feature parameter extractor.

When registering a word using the device shown in FIG. The audio feature parameters are manually extracted and stored and registered in the memory 2.

At this time, character codes representing the words are input from the keyboard 12 and are stored and registered in the memory 2.

Note that these control QUs are performed by a processor 7, and 13 indicates a bus.

Next, the voice feature parameter extractor will be explained. The one currently used is one that obtains a spectral envelope using a band pass as shown in FIG.

This lasts about 2 seconds (word speech is converted into a digital signal by the A/D converter 6, and one frame has about 512 samples, and the passband width of about 5 to 30 samples per frame is 100 samples). ~200Hz bandpass filter group 20
Therefore, the low range ranges from 100Hz to the high range of about 7000Hz (3400Hz for telephone voice).
Spectral analysis is performed, normalized by a normalization group 21, and analyzed values passed through a low-pass filter group 22 with a cutoff frequency of several Hz to Q10 Hz are integrated to obtain a spectral envelope, which is a voice characteristic parameter. be.

[Problem that the invention seeks to solve]

However, in the conventional word registration method, the audio feature parameters of the input word are extracted and then stored and registered in memory, so words with similar characteristics may be stored and registered. There is a problem that there are many erroneous recognitions during recognition.

[Means for solving problems]

The above problem is solved by the speech feature parameter extractor 1 of the word uttered for registration, as shown in the principle block diagram of Fig. 1.
The speech feature parameters extracted in the above are compared with the speech feature parameters of the words already stored and registered in the memory 2 in the speech feature parameter comparison section 3, and if there are any words with similar degrees of similarity, the differences between the similar words are determined. This problem is solved by the word registration method of the present invention, in which the word registration method of the present invention is performed by asking the person to say the word with emphasis, or by having the person say a word with the same meaning in a different way, and then re-registering the word.

[Effect]

According to the present invention, when registering a word, the phonetic feature parameters of this word are compared with the phonetic feature parameters of words that have already been stored and registered in the memory 2, and if there are words with similar degrees of similarity,
For example, by asking people to say words that have the same meaning, such as nicknames, and registering them again, we avoid registering words with a high degree of similarity. It can be reduced.

〔Example〕

An embodiment of the present invention will be described below with reference to the accompanying drawings.

FIG. 2 is a block diagram of a word registration device according to an embodiment of the present invention.

To explain the word registration in FIG.
The input signal is manually inputted, converted into a digital signal, and input to the voice feature parameter extractor 1, where the voice feature parameters are extracted and input to the voice feature parameter comparison unit 3 using a dynamic programming matching method or the like.

On the other hand, the voice feature parameters already stored and registered in the memory 2 are sequentially read out under the control of the processor 7, sent to the voice feature parameter comparison section 3, and compared sequentially. At the same time, the stored and registered character code is displayed on the display section 8, and this character code is also sent to the voice synthesis section 9, where a voice signal is synthesized, and the word voice is outputted from the speaker 11.

Then, the specific speaker can see that the word he just uttered is similar to the voice feature parameters already stored and registered in memory 2, so the specific speaker may be asked to rephrase it in a clearer way, or give him a nickname or other similar meaning. Since the user is trying to buy by saying the word he or she has, it is possible to reduce misrecognition during word speech recognition.

In this case as well, these controls are performed by the processor 7.

Note that the voice feature parameters stored and registered in this memory 2 are used during word voice recognition.

Next, a dynamic programming matching method (DP matching method) currently used as a method for comparing voice feature parameters will be explained.

This is done by associating the time axes so as to obtain the best matching between the input word speech feature parameters and the word speech feature parameters being compared, and to eliminate the effects of uneven expansion and contraction of the time axis in word speech. Perform time normalization without receiving any data, then compare each point of both patterns to find the time normalized distance, calculate this value for each pattern, find the sum, and if this sum is less than a predetermined value It is assumed that the degree of similarity is obtained.

〔Effect of the invention〕

As explained in detail above, (according to the present invention, if there is a word that has a similar degree of similarity to the voice feature parameters of a word that has already been stored and registered in the memory, the difference between the word and the similar word is emphasized and the voice is uttered.) , or a word that has the same meaning in a different way, and then registers it again. This prevents the registration of words that have a similar degree of similarity to the voice feature parameters of words that have already been stored and registered in memory. This has the effect of improving the recognition rate during recognition.

[Brief explanation of the drawing]

Fig. 1 is a block diagram of the principle of the present invention, Fig. 2 is a block diagram of a word registration device according to an embodiment of the present invention, Fig. 3 is a block diagram of a conventional word registration device, and Fig. 4 is an example of speech. FIG. 2 is a block diagram of a feature parameter extractor. In the figure, 1 is a voice feature parameter extractor, 2 is a memory, 3 is a voice feature parameter comparison unit, 5 is a microphone, 6 is an A/D conversion unit, 7 is a processor, 8 is a display unit, 9 is a voice synthesis unit, 11 is the speaker, Tsumeda Tagawa's big play
Furohei 2 Kuchiju 4-9's Go-q Month) Sen 9 1 vs. niece's FQ So 2 Thirst
3 Ri

Claims (1)

[Claims] When registering words of a specific speaker for word speech recognition, the speech feature parameters extracted by the speech feature parameter extractor (1) of the words uttered for registration are already stored in memory. The voice feature parameters of the words stored and registered in (2) are compared in the voice feature parameter comparison section (3), and if there are words with a similar degree of similarity, the words are uttered while emphasizing the differences from similar words. A word registration method characterized in that the word registration is performed again after receiving the word or having the word uttered in a different way but having the same meaning.