HUMAN SPEECH RECOGNITION APPARATUS AND METHOD

• US 20090157403A1
• Filed: 12/12/2008
• Published: 06/18/2009
• Est. Priority Date: 12/14/2007
• Status: Active Grant

• Current Assignee: Electronics and Telecommunications Research Institute
• Original Assignee: Electronics and Telecommunications Research Institute
• Inventors: Lee, Yunkeun, CHUNG, Hoon
• Granted Patent: US 8,185,393 B2
• Time in Patent Office: Days
• US Class Current: 704/254
• CPC Class Codes: G10L 15/187 Phonemic context, e.g. pronunciation rules, phonotactical constraints or phoneme n-grams
• CPC Class Codes: G10L 2015/025 Phonemes, fenemes or fenones being the recognition units

Abstract

A speech recognition apparatus generates a feature vector series corresponding to a speech signal, and recognizes a phoneme series corresponding to the feature vector series using sounds corresponding to phonemes and a phoneme language model. In addition, the speech recognition apparatus recognizes vocabulary that corresponds to the recognized phoneme series. At this time, the phoneme language model represents connection relationships between the phonemes, and is modeled according to time-variant characteristics of the phonemes.

Claims

1. A speech recognition method that recognizes vocabulary from a speech signal, the method comprising:

generating a feature vector series that corresponds to the speech signal;

recognizing a phoneme series corresponding to the feature vector series using a phoneme language model, which indicates connection relationships between phonemes and is modeled in consideration of a location where each of the phonemes is disposed in arbitrary vocabulary; and

recognizing vocabulary that corresponds to the phoneme series.

2. The speech recognition method of claim 1, wherein the phoneme language model is the probability in the case in which one of a plurality of phonemes recognized at an arbitrary time is recognized in a state where previous (n−

1) phonemes are recognized, with respect to each of the plurality of phonemes.

3. The speech recognition method of claim 2, wherein the recognition of the phoneme series includes recognizing the phoneme series using a phoneme condition, which is the probability in the case of observing a feature vector series derived from each of the plurality of phonemes.

4. The speech recognition method of claim 3, wherein the recognition of the phoneme series further includes recognizing a phoneme series having the maximum posterior probability with respect to the feature vector series on the basis of the phoneme condition and the phoneme language model.

5. The speech recognition method of claim 2, wherein the recognizing of the vocabulary includes searching a comparison vocabulary having the maximum posterior probability with respect to the phoneme series on the basis of a vocabulary condition indicating a frequency of the arbitrary vocabulary, a comparison phoneme series obtained by subjecting the arbitrary vocabulary to pronunciation conversion, and an edition distance.

6. The speech recognition method of claim 5, wherein the recognizing of the vocabulary further includes recognizing the edition distance in a search space by each of phonemes of the comparison phoneme series and each of phonemes of the phoneme series.

7. The speech recognition method of claim 6, wherein the recognizing of the vocabulary includes:

dividing the comparison phoneme series into a plurality of continuous connection phonemes;

dividing the phoneme series into a plurality of continuous connection phonemes; and

recognizing the edition distance in a search space by each of the connection phonemes of the comparison phoneme series and each of the connection phonemes of the phoneme series.

8. The speech recognition method of claim 7, wherein the plurality of connection phonemes include at least one vowel.

9. A speech recognition apparatus that recognizes vocabulary from a speech signal, the apparatus comprising:

a feature vector series generator that generates a feature vector series corresponding to the speech signal;

a phoneme recognition unit that derives a phoneme series corresponding to the feature vector series using a phoneme language model, which indicates connection relationships between phonemes and is modeled in consideration of a location where each of the phonemes is disposed in arbitrary vocabulary; and

a vocabulary recognition unit that derives vocabulary that corresponds to the phoneme series.

10. The speech recognition apparatus of claim 9, wherein the phoneme language model is the probability in a case in which one of a plurality of phonemes recognized at an arbitrary time is recognized in a state where previous (n−

1) phonemes are recognized, with respect to each of the plurality of phonemes.

11. The speech recognition apparatus of claim 10, wherein the phoneme recognition unit recognizes a phoneme series having the maximum posterior probability with respect to the feature vector series on the basis of the probability in the case of observing an arbitrary feature vector series from each of the plurality of phonemes and the phoneme language model.

12. The speech recognition apparatus of claim 10, wherein the vocabulary recognition unit derives vocabulary having the maximum posterior probability with respect to the phoneme series on the basis of a vocabulary condition indicating a frequency of the arbitrary vocabulary, a comparison phoneme series obtained by subjecting the arbitrary vocabulary to pronunciation conversion, and the edition distance of the comparison phoneme series and the phoneme series.

13. The speech recognition apparatus of claim 12, wherein the vocabulary recognition unit recognizes the edition distance in a search space by each of phonemes of the comparison phoneme series and each of phonemes of the phoneme series.

14. The speech recognition apparatus of claim 13, wherein the vocabulary recognition unit recognizes the edition distance in a search space by each of a plurality of connection phonemes obtained by dividing the comparison phone series, and each of a plurality of connection phonemes obtained by dividing the phoneme series.

15. The speech recognition apparatus of claim 14, wherein the plurality of connection phonemes include at least one vowel.