JPH0119599B2 - - Google Patents

Description

[Detailed Description of the Invention] (a) Technical Field of the Invention The present invention relates to a speech recognition device, and particularly to monosyllabic speech recognition that recognizes unknown input monosyllabic speech of a speaker whose monosyllabic speech has been registered in the speech recognition device in advance. Regarding the method.

(b) Background of the Technology As speech recognition technology has improved in recent years, there has been a desire for a speech recognition device with fewer recognition errors when recognizing a speaker's speech. The speech recognition method mainly converts the speaker's monosyllabic speech into feature parameters and stores them in advance, and then compares the feature parameters of the unknown input monosyllabic speech with the pre-stored feature parameters to find the one that is most similar. It is recognized as a corresponding monosyllabic voice, but even if the same monosyllabic voice is pronounced, the characteristic parameters change depending on the way it is uttered, so even if you register the same monosyllabic voice several times with different utterance methods, errors will occur. It is difficult to reduce to zero. In particular, for monosyllabic speech that has characteristic parameters that are likely to cause recognition errors, it is impossible to improve the recognition rate unless the matching method is taken into consideration. For this reason, after comparing all monosyllabic voices registered in advance with the monosyllabic voice of the speaker, multiple replays are performed in order from the monosyllabic voice that is most similar to the unknown input monosyllabic voice based on the matching results. A matching candidate is selected from registered monosyllabic speech, and the unknown input monosyllabic speech is re-matched with the plurality of re-matching candidates using a re-matching parameter determined according to the combination of the plurality of re-matching candidates to improve the recognition rate. A monosyllabic speech recognition method has been proposed. However, there is room for improvement in the above reverification method, and countermeasures are desired.

(c) Purpose of the Invention Based on the above-mentioned needs, the purpose of the present invention is to narrow down the number of re-verification candidates, shorten the time required for re-verification, and improve the configuration of the device in the monosyllabic speech recognition method using the re-verification method. The aim is to simplify and improve economic efficiency.

(d) Structure of the Invention The structure of the present invention is to register monosyllabic speech in advance, and to calculate the characteristic parameters of the unknown input monosyllabic speech and the characteristic parameters of all monosyllabic speech registered in advance.
After DP matching, multiple rematching candidates are selected from the registered monosyllabic speech in descending order of similarity, and unknown input monosyllables are determined by rematching parameters determined according to the combination of the multiple rematching candidates. The speech is re-matched with the re-matching candidates, and the re-matching candidate that is most similar as a result is recognized as the corresponding monosyllabic speech, but when selecting the multiple re-matching candidates,
If the re-matching candidate with the highest degree of similarity at the time of DP matching is the one with the highest degree of similarity, or the combination of multiple re-matching candidates with the highest degree of similarity is predetermined, the re-matching step is omitted and the DP matching is performed. The rematching candidate with the highest similarity in time is sent out as the recognition result, thereby shortening the monosyllabic speech recognition time and simplifying the rematching circuit.

(e) Embodiment of the invention The figure is a block diagram of a circuit showing an embodiment of the invention. First, the speaker connects the switching section 3 to the parameter storage section 4 under the control of the control section 8 in order to register monosyllabic speech in advance, and inputs the monosyllabic speech.
The preprocessing section 1 performs audio level adjustment, analog-to-digital conversion, etc., and sends the result to the parameter extraction section 2. The parameter extraction section 2 extracts characteristic parameters of the monosyllabic speech and stores them in the parameter storage section 4. Next, in order to recognize the monosyllabic speech, the speaker connects the switching section 3 to the storage section 5 under the control of the control section 8, and utters the monosyllabic speech. The characteristic parameters of the unknown input monosyllabic speech that have entered the storage unit 5 via the preprocessing unit 1, parameter extraction unit 2, and switching unit 3 through the same operation as described above are stored in the parameter storage unit 4 under the control of the control unit 8. The feature parameters of all monosyllabic voices are compared with the DP in the matching unit 6, and the monosyllabic voice with the most similar feature parameters among the feature parameters of all monosyllabic voices is selected as the first candidate for re-matching. Then, a plurality of re-verification candidates are sequentially selected and sent to the determination unit 7. The determination unit 7 compares the first re-verification candidate or the combination of the top re-verification candidates with the candidates stored in advance in the table, and if the same candidates exist, the re-verification step is omitted and control is performed. After passing through section 8, the candidate selected as the first re-verification candidate by the collation section 6 is selected as a recognition result. If the first re-verification candidate selected by the collation unit 6 or a combination of the top re-verification candidates is other than a predetermined combination, it is sent from the determination unit 7 to the control unit 8 for re-verification and recognition. The control unit 8 causes the parameter storage unit 4 to send the feature parameters corresponding to the re-verification candidate to the multiplier 10 and the feature parameters of the unknown input monosyllabic speech stored in the storage unit 5 to the multiplier 11, respectively, and sends the feature parameters corresponding to the re-verification candidate to the multiplier 11, 7 multiplies from the frequency weight storage unit 12 a rematching parameter determined by the rematching candidate, that is, a parameter that emphasizes frequency band components suitable for mutually identifying rematching candidates and reduces other frequency band components. Vessels 10, 11
send to. Further, the determination unit 7 causes the threshold value storage unit 13 to send a threshold value, which is a parameter for determining the optimal matching interval determined according to the re-matching candidate, to the re-matching unit 9. The re-verification unit 9 performs re-verification using the outputs of the multipliers 10 and 11 and the threshold value from the threshold storage unit 13. A plurality of re-verification candidates are re-verified with the unknown input monosyllabic speech in order from the first candidate, and the most similar re-verification candidate is outputted from the control unit 8 as a recognition result.

In this embodiment, examples of candidates for omitting the re-verification process stored in the predetermined table of the determination unit 7 are: those with extremely high recognition rates in the verification unit 6;
That is, for example, the first rematching candidate is as follows: a monosyllabic voice with a consonant beginning with /w/, such as wa, and a monosyllabic voice with a consonant beginning with /j/, such as yuyo. Combinations that are unlikely to appear as re-verification candidate combinations, such as the following.

Monosyllabic consonants starting with /b/ and /t/ as in ba and ta Monosyllabic consonants starting with /d/ and /k/ as in da and ka Monosyllabic consonants starting with /d/ and /k/ as in ma and ha consonants /m/ and / Monosyllabic sounds starting with h/ These are monosyllabic sounds where the consonants begin with /g/ and /p/, such as ga and ba.

/b/ and /t/, /d/ and /k/, /m/
, /h/, /g/, and /p/ are less likely to be mistaken for each other, and the results of the matching unit 6 are used because the monosyllabic sounds that form these combinations are regarded as utterances that do not require re-verification. .

Monosyllabic sounds where the consonant starts with /r/, such as A, where the effect of rematching is small.
r/ is unstable and has large variations, and tends to be mistaken for various other monosyllabic sounds, so re-verification has little effect.Currently, cost performance is poor and it is more advantageous to omit re-verification.

(f) Effects of the Invention As explained above, the present invention reduces the time required for re-verification by narrowing down the number of re-verification candidates in a monosyllabic speech recognition method using a re-verification method, and also improves the performance related to the re-verification operation. This has a great effect because it is possible to simplify the component equipment used in the process and improve economic efficiency.

[Brief explanation of drawings]

The figure is a block diagram of a circuit showing one embodiment of the present invention. 1 is a preprocessing unit, 2 is a parameter extraction unit,
3 is a switching unit, 4 is a parameter storage unit, 5 is a storage unit, 6 is a verification unit, 7 is a determination unit, 8 is a control unit, 9 is a re-verification unit, 10 and 11 are multipliers, 12 is a frequency weight storage unit , 13 is a threshold storage unit.

Claims (1)

[Claims] 1. After comparing all previously registered monosyllabic speech with the unknown input monosyllabic speech, multiple re-matching candidates are selected from the registered monosyllabic speech based on the matching results, and a combination of the re-matching candidates is performed. In a speech recognition device that selects rematching parameters for each case and rematches unknown input monosyllabic speech, when selecting the rematching candidates, the rematching candidate with the highest degree of similarity is determined in advance. 1. A monosyllabic speech recognition method characterized in that re-verification is omitted when a combination of re-verification candidates with high similarity is determined in advance.