JPH0331275B2 - - Google Patents

Description

[Detailed Description of the Invention] (A) Technical Field of the Invention The present invention relates to a speech standard feature pattern creation processing device,
In particular, it is possible to easily delete registration of standard feature patterns caused by incorrect utterances by playing back and letting the user listen to registered sounds related to standard feature patterns that are used for matching with input feature patterns obtained from unknown input sounds. The present invention relates to a speech standard feature pattern creation processing device.

(B) Prior art and problems In general, in speech recognition, in order to improve the recognition rate, it is important to determine what kind of feature parameters are extracted from speech information and used for matching. It is also important to determine how best to prepare the standard feature parameters representative of each item in the dictionary through extraction. No matter how good the method of feature extraction or matching is, the standard feature patterns registered in the dictionary may contain defective standard feature patterns such as noise-added patterns, unclear utterance patterns, or, for example, "a". If there are many erroneous standard feature patterns due to utterance errors such as uttering ``i'' when ``i'' should be registered, the recognition rate will not improve.

Standard feature patterns are stored as digital information in the dictionary, and there are many of them, and they are not visible like mechanical parts, and not all standard feature patterns are uniformly used, so once registered. If this occurs, it is not easy to detect the above-mentioned defective standard feature patterns, erroneous standard feature patterns, etc.

Conventionally, all standard feature patterns once registered are treated as correct, and if a recognition error occurs,
Generally, this was unavoidable because the input voice to be recognized was bad or at the limits of recognition. In addition, a learning method has been proposed that improves the quality of the dictionary by averaging the input feature pattern extracted from the input voice that caused the misrecognition with the already registered standard feature pattern. However, this method is based on the premise that the registered standard feature pattern is correct to some extent, and if the standard feature pattern is incorrect, there is a problem that convergence is slow.

If a standard feature pattern to be registered or a standard feature pattern that has already been registered is not appropriate, it is desirable to be able to detect it and re-register it.

By the way, in the past, methods have been considered to output voice recognition results as voice, but the voice information output as recognition results is prepared corresponding to each item, and is based on each standard feature. It was not prepared in accordance with the pattern. Therefore, in the case where a recognition error occurs, it is not possible to determine whether the standard feature pattern is good or bad even if the output audio is heard.

(C) Object and Structure of the Invention The present invention aims to solve the above-mentioned problems, and allows the user to The aim is to improve the quality of the dictionary and increase the recognition rate by making it possible to recognize that the dictionary has been registered as an incorrect utterance, and by re-registering invalid standard feature patterns. Therefore, the speech standard feature pattern creation processing device of the present invention compares an input feature pattern obtained by acoustically analyzing unknown input speech with a standard feature pattern stored in advance corresponding to each item in a dictionary. A speech standard feature pattern creation processing device in a speech recognition system that performs speech recognition, wherein the dictionary includes a speech information storage section that stores speech information corresponding to the standard feature pattern, an audio playback unit that reproduces and outputs audio based on the corresponding audio information in the dictionary when this occurs, and deletes the registration of the standard feature pattern selected at the time of the recognition from the dictionary in response to an instruction to delete the registration. It is characterized by having a registration/deletion section to do the following. Referring to the drawings below,
An explanation will be given according to an example.

(D) Embodiments of the invention FIG. 1 is a diagram for explaining the relationship between the distribution of voice patterns and standard feature patterns, FIG. 2 is a diagram for explaining the outline of the processing according to the present invention, and FIG. As an example of reference technology related to the embodiment, FIG. 4 shows an explanatory diagram of audio reproduction in the embodiment illustrated in FIG. 3.

In Figure 1, the parts surrounded by solid lines A, B, and C indicate the distribution of actual speech patterns in the pattern space, and A ₁ and A ₂ are for word A (including monosyllables; the same applies hereinafter). The registered standard feature patterns B ₁ to B ₃ represent standard feature patterns for word B, and C ₁ represents a standard feature pattern for word C. In some cases, one word item can be covered by one standard feature pattern, as shown in illustration C, but usually, multiple standard feature patterns are prepared and recognized for one word, as shown in illustrations A and B. It usually covers the distribution range of power speech patterns. For example, when an input feature pattern X of unknown input speech is extracted, the input feature pattern X and each standard feature pattern A ₁ , A ₂ ,
The matching distance with B ₁ is calculated, and the item to which the standard feature pattern with the short distance belongs is taken as the recognition result.

If the standard feature patterns registered in the dictionary include a defective standard feature pattern or an erroneous standard feature pattern that deviates from the distribution of speech patterns, the recognition rate will deteriorate. The present invention aims to improve the recognition rate by deleting such invalid standard feature patterns after registration.

For example, as shown in FIG. 2, it is assumed that the distribution of the voice pattern of the word "Shibuya" is S in the diagram, and the distribution of the voice pattern of the word "Hibiya" is H in the diagram. When creating a dictionary, when registering multiple standard feature patterns, due to a mistake in operation or pronunciation, the user mistakenly pronounced "hibiya" instead of "shibuya", resulting in the standard feature pattern being registered. Assume that feature pattern _S3 has been registered. Standard feature pattern S ₃ is stored in the dictionary as belonging to the word "Shibuya" even though it is actually a speech pattern for "Hibiya".

Once registered as described above, for example, when recognizing the utterance of "Jibuya",
Only standard feature patterns S ₁ and S ₂ match, and pattern S ₃ does not match. However, the fact that pattern _S3 is incorrect is not detected, and it is simply treated as if no utterance corresponding to pattern _S3 is made. On the other hand, it is assumed that the input feature pattern X for "Hibiya" is uttered as shown in FIG. 2, for example. Since the distance d ₁ between the input feature pattern X and the standard feature pattern S ₃ is smaller than the distance d ₂ between the input feature pattern X and the standard feature pattern H ₃ , the pattern X is recognized as the word "Shibuya". In this case, according to conventional learning methods, it is determined that the standard feature patterns H ₁ , H ₂ , H ₃ of the word "Hibiya" are inappropriate, rather than that the standard feature pattern S ₃ is incorrect. Standard feature patterns belonging to "Hibiya" were added and modified. Therefore, the error standard feature pattern _S3 will be left as is in the dictionary.

In the case of the reference technology shown in Figure 3, by playing back and outputting the audio input when attempting to register a standard feature pattern on the spot, it is possible to check whether the output audio is the same as the one intended for registration. The registration of error standard feature patterns such as the standard feature pattern _S3 above is
prevent it from happening. Furthermore, the present invention
As will be described later with reference to FIG. 5, even if the error standard feature pattern _S3 is registered by mistake, when the standard _feature pattern _S3 is used for recognition candidate, By playing and outputting the registered audio, the standard feature pattern S ₃
to be able to judge whether or not it is appropriate. This makes it possible to detect defective standard feature patterns and eliminate them from the dictionary.

FIG. 3 is a block diagram showing reference technology related to this embodiment, and in the figure, reference numeral 1 is a microphone, 2 is an acoustic analysis section, 3 is a pattern extraction section, 4 is a switching section, and 5 is an audio reproduction section. 6 is a speaker, 7 is an input pattern buffer, 8 is a selection key, 9 is an error indication section, 10 is a pattern addition section, 11 is a registration rejection section, 12 is a dictionary, and 13 is a collation determination section.

The audio signal input from the microphone 1 is subjected to frequency analysis in the acoustic analysis section 2. The acoustic analysis unit 2 includes, for example, a group of band filters, a parameter extraction circuit, etc., and extracts features (parameters) of the input speech, such as a moment _M1 corresponding to the first formant frequency and a moment M1 corresponding to the second formant frequency. Moment _M2 , low frequency power, high frequency power, etc. are extracted, sample points related to these feature quantities are determined, and time series information of the feature quantities is obtained.

The parameter time series information obtained by the acoustic analysis section 2 is input to the pattern extraction section 3. The pattern extraction unit 3 extracts an input feature pattern representing the characteristics of the input voice from this parameter time series information. The switching unit 4 switches registration or verification of pattern information in response to a mode switching instruction from, for example, a keyboard (not shown).

When the registration mode is designated, the input feature pattern is held in the input pattern buffer 7. The audio reproduction unit 5 reproduces the input audio for registration by voice synthesis or the like and outputs it from the speaker 6. With this, for example, if you accidentally say "Hibiya" when you should say "Shibuya", you can confirm the utterance with your ears before registering.
Errors can be detected. The selection key 8 is a key for selecting whether to register or not. The error instruction unit 9 activates the pattern addition unit 10 in the case of registration, and activates the registration rejection unit 11 in the case of not registering due to a utterance error. Pattern addition part 10
uses the input feature pattern stored in the input pattern buffer as a standard feature pattern in dictionary 1.
Add registration to 2. The registration rejection unit 11 rejects the input feature pattern on the input pattern buffer 7 and prompts for re-voice input. The matching/judgment unit 13 matches the input feature pattern with the standard feature pattern in the dictionary in the recognition mode, and outputs the recognition result.

In addition, in FIG. 3, the pattern adding section 10
Alternatively, the pattern may be additionally registered unconditionally, and the registration rejecting unit 11 may later cancel the registration based on an instruction from the error indicating unit 9.

The audio reproduction section 5 can reproduce input audio and output it to the speaker 6, for example, as shown in FIG. In the case shown in FIG. 4A, the voice is output by synthesizing the voice based on the feature parameter time series obtained by acoustically analyzing the registered voice. Note that this speech synthesis can be realized by a well-known technique, so the explanation will be limited to this extent here.

In the case shown in FIG. 4B, the digital audio obtained by analog/digital conversion of the registered audio is stored as it is in the audio buffer, and the audio reproducing section performs the digital/analog conversion and outputs it. In the case shown in Figure 4 C, the digital audio obtained by converting the registered audio from analog to digital is encoded, stored in an audio buffer, and then decoded and converted from digital to analog to reproduce the audio. .

FIG. 5 shows the configuration of an embodiment of the present invention, and FIG.
An explanatory diagram of the audio reproduction method of the illustrated embodiment is shown. In the figure, numerals 1 to 6, 12, and 13 correspond to those in FIG. 3, 20 is a pattern/voice registration section, 21 is an erroneous recognition instruction key, 22 is a registration deletion instruction key, 23 is an erroneous recognition instruction section, and 24 is an erroneous recognition instruction key. The registration deletion unit 25 represents a voice information storage unit.

In the embodiment illustrated in FIG. 3, the registered voice is played back and output at the time of registration, but in the embodiment illustrated in FIG.
The registered audio related to the selected standard feature pattern is played back and output. Therefore, even after registration, it is possible to detect and re-register the erroneous standard feature pattern.

Acoustic analysis section 2, pattern extraction section 3, switching section 4
is similar to that explained in FIG. In the registration mode, the pattern/voice registration unit 20 sends voice information, which will be described later with reference to FIG. and register.

In the recognition mode, the matching determination unit 13 calculates the distance between the input feature pattern and each standard feature pattern,
Perform verification. Then, the one with the smallest distance is output as the recognition result. The user presses the erroneous recognition instruction key 21 when the recognition result is incorrect. The erroneous recognition instruction section 23 detects the depression of the erroneous recognition instruction key 21 and notifies the processing section of the dictionary 12 or the audio reproduction section 5 of the recognition error. The audio reproduction unit 5 reads audio information corresponding to the first recognition candidate or a standard feature pattern within a predetermined range or within a user's requested range from the audio information storage unit 25, reproduces the audio, and outputs the audio to the speaker 6. Output from.

Common causes of misrecognition include cases where the unknown input voice is unclear and cases where the standard feature pattern is inappropriate. In the case of the present invention, the validity of the standard feature pattern selected at the time of recognition can be checked by the audio output by the audio reproduction section 5. If the standard feature pattern is not valid, the user presses the registration deletion instruction key 22. By pressing this button, the registration deletion unit 24 deletes the standard feature pattern and its audio information from the dictionary 12. Then, if necessary, switch to registration mode and re-register the correct standard feature pattern. Note that, as described above, the audio reproducing unit 5 may reproduce and output the audio only when a recognition error occurs, or when the matching determination unit 13 outputs the recognition result, it may output the audio based on the recognition result. Therefore, the voice may be output regardless of the presence or absence of a recognition error.

Next, the audio reproduction process will be explained with reference to FIG. For example, as shown in FIG. 6A, at the time of registration, the characteristic parameter time series obtained by acoustically analyzing the registered speech is stored in the dictionary 12. The audio reproduction unit 5 reads out the characteristic parameter time series,
The audio is synthesized and output to the speaker 6. In addition, as shown in FIG. 6B, the digital voice obtained by converting the registered sound from analog to digital is stored in the dictionary 12 as it is, read out at the time of recognition, performs digital/analog conversion, and reproduces the voice. It's okay. Furthermore, as shown in FIG. 6C, in order to reduce the storage area of the dictionary 12, the registered speech is encoded after analog/digital conversion, the encoded information is stored in the dictionary 12, and the encoded information is decoded during recognition and played back. It may also be possible to output it.

(E) Effects of the Invention As explained above, according to the present invention, it becomes possible to easily delete the registration of defective feature patterns such as noise-added patterns and unclear speech patterns, as well as erroneous standard feature patterns due to incorrect speech production. It is possible to improve the quality of the dictionary and improve the recognition rate.

[Brief explanation of the drawing]

Fig. 1 is a diagram for explaining the relationship between the distribution of voice patterns and standard feature patterns, Fig. 2 is a diagram for explaining the outline of the processing according to the present invention, and Fig. 3 is a reference technique related to this embodiment. example, FIG. 4 is an explanatory diagram of audio reproduction in the embodiment illustrated in FIG. 3,
FIG. 5 shows the configuration of an embodiment of the present invention, and FIG. 6 is an explanatory diagram of the audio reproduction system of the embodiment shown in FIG. In the figure, 3 is a pattern extraction section, 5 is an audio reproduction section,
9 is an error instruction section, 11 is a registration rejection section, 12 is a dictionary, 13 is a collation determination section, 23 is an error recognition instruction section, 2
4 represents a registration deletion unit, and 25 represents an audio information storage unit.

Claims (1)

[Claims] 1. Speech recognition is performed by comparing an input feature pattern obtained by acoustic analysis of unknown input speech with a standard feature pattern stored in advance corresponding to each item in a dictionary. A voice standard feature pattern creation processing device in a recognition system, wherein the dictionary includes a voice information storage unit that stores voice information corresponding to the standard feature pattern, and the dictionary is used during voice recognition or when a recognition error occurs. an audio reproduction unit that reproduces and outputs audio based on the corresponding audio information in the dictionary; and a registration deletion unit that deletes the registration of the standard feature pattern selected at the time of recognition from the dictionary in response to a registration deletion instruction. 1. A speech standard feature pattern creation processing device, characterized in that: 2. The voice information stored in the dictionary is feature parameter time series extracted from registered voices, digital information obtained by A/D conversion of registered voices, or audio encoded information thereof. The speech standard feature pattern creation processing device according to item 1.