Method and apparatus for tone-sensitive acoustic modeling

• US 5,884,261 A
• Filed: 07/07/1994
• Issued: 03/16/1999
• Est. Priority Date: 07/07/1994
• Status: Expired due to Term

• Current Assignee: Apple Inc.
• Original Assignee: Apple Computer Incorporated (Apple Inc.)
• Inventors: Hon, Hsiao-Wuen, Yuan, Baosheng, de Souza, Peter V., Fineberg, Adam B.
• Primary Examiner(s): Dorvil, Richemond
• Application Number:
  US08/271,639
• Time in Patent Office: 1,713 Days
• Field of Search: 395/2.65, 395/2.52, 395/2.16, 395/2.64, 395/2.49, 395/2.63, 381/41-43, 704/256, 704/243, 704/255, 704/254, 704/240, 704/205, 704/206, 704/207, 704/208, 704/209, 704/200, 704/241, 704/242, 704/216, 704/217, 704/218, 704/263, 704/265, 704/266 , 704/268, 704/267, 704/269
• US Class Current: 704/255
• CPC Class Codes: G10L 15/144 Training of HMMs
• CPC Class Codes: G10L 2015/025 Phonemes, fenemes or fenones being the recognition units
• CPC Class Codes: G10L 25/15 the extracted parameters being formant information
• CPC Class Codes: G10L 25/90 Pitch determination of speech signals

Abstract

Tone-sensitive acoustic models are generated by first generating acoustic vectors which represent the input data. The input data is separated into multiple frames and an acoustic vector is generated for each frame which represents the input data over its corresponding frame. A tone-sensitive parameter is then generated for each of the frames which indicates the tone of the input data at its corresponding frame. Tone-sensitive parameters are generated in accordance with two embodiments. First, a pitch detector may be used to calculate a pitch for each of the frames. If a pitch cannot be detected for a particular frame, then a pitch is created for that frame based on the pitch values of surrounding frames. Second, the cross covariance between the autocorrelation coefficients for each frame and its successive frame may be generated and used as the tone-sensitive parameter. Feature vectors are then created for each frame by appending the tone-sensitive parameter for a frame to the acoustic vector for the same frame. Then, using these feature vectors, acoustic models are created which represent the input data.

Claims

1. A method for generating a tone-dependent acoustic model comprising the steps of:

(a) generating a plurality of acoustic vectors representing a plurality of input data frames;

(b) generating a plurality of tone-sensitive parameters corresponding to said data frames;

(c) creating a plurality of feature vectors corresponding to said data frames by appending said plurality of tone-sensitive parameters to said plurality of acoustic vectors; and

(d) creating an acoustic model from said plurality of feature vectors.

2. The method as defined in claim 1 wherein each of said data frames corresponds to a single vector of said feature vectors.

3. The method as defined in claim 1 wherein said data frames represent sampled frames of speech.

4. The method as defined in claim 1, wherein each of said tone-sensitive parameters corresponds to a single frame of said input data frames.

5. The method as defined in claim 1, wherein said step (b) comprises detecting pitch values.

6. The method as defined in claim 5, wherein said step (b) further comprises the steps of:

detecting whether a pitch value is present in each of said data frames; and

determining a pitch value for each of said input data frames.

7. The method as defined in claim 6, wherein for a first frame of said data frames, said step of determining pitch values further comprises the steps of:

using said pitch value detected for said first frame, provided said pitch value is present in said first frame; and

generating a pitch value for said first frame based on detected pitch values of surrounding frames, provided said pitch is not present for said first frame.

8. The method as defined in claim 6, wherein said step (c) comprises appending said pitch values to said acoustic vectors for said data frames.

9. The method as defined in claim 1, wherein said step (b) comprises the steps of:

determining autocorrelation coefficients for each of said data frames; and

determining cross covariances between the autocorrelation coefficients of adjacent data frames.

10. The method as defined in claim 9, wherein said step (c) comprises appending said cross covariances to said acoustic vectors.

11. The method as defined in claim 10, wherein said step (b) further comprises the steps of:

detecting whether a pitch value is present in each of said data frames; and

determining a pitch value for each of said input data frames.

12. The method as defined in claim 11, wherein said step (c) further comprises appending said pitch values to said acoustic vectors for said data frames.

13. The method as defined in claim 1, wherein said step (d) comprises the steps of:

creating training vectors based on said feature vectors; and

constructing acoustic models using said training vectors.

14. The method as defined in claim 13 wherein said creating step comprises the steps of:

concatenating each of said feature vectors with surrounding feature vectors to generate a giant vector corresponding to each of said feature vectors;

associating the giant vector with an aligned arc corresponding to said feature vector;

computing eigenvectors which optimally discriminate between said arcs; and

calculating the dot product of the eigenvectors and the vector representing the input data frame.

15. A method generating tone-sensitive vectors representing frames of input data from which acoustic models can be created, said method comprising the steps of:

(a) creating an acoustic vector for a first frame of said input data;

(b) determining a pitch value for said first frame; and

(c) appending said pitch value to said acoustic vector for said first frame.

16. The method as defined in claim 15, wherein said determining step comprises:

using the pitch value detected by a pitch detector for said pitch value for said first frame, provided said pitch value is present in said first frame; and

generating a pitch value for said first frame based on detected pitch values of surrounding frames, provided said pitch is not present for said first frame.

17. A method of generating a tone-sensitive feature vector for use in a speech recognition system representing a first frame of input data, wherein said input data comprises multiple frames, said method comprising the steps of:

(a) determining autocorrelation coefficients for each of said multiple frames;

(b) determining a cross covariance between the autocorrelation coefficients for said first frame and a second frame of said multiple frames; and

(c) appending said cross covariance to an acoustic vector for said first frame.

18. The method as defined in claim 17 wherein said step (a) comprises determining autocorrelation coefficients for a predetermined number of lags.

19. The method as defined in claim 17 wherein said step (b) comprises determining said cross covariance by shifting said second frame relative to said first frame a predetermined amount.

20. A method of generating tone-dependent acoustic models comprising the steps of:

(a) generating feature vectors representing input data frames, wherein each of the feature vectors includes a tone-sensitive parameter;

(b) creating training vectors based on said feature vectors; and

(c) constructing acoustic models using said training vectors.

21. The method as defined in claim 20 wherein said creating step comprises the steps of:

concatenating each of said feature vectors with surrounding feature vectors to generate a giant vector corresponding to each of said feature vectors;

associating the giant vector with an aligned arc corresponding to said feature vector;

computing eigenvectors which optimally discriminate between said arcs; and

calculating the dot product of the eigenvectors and the vector representing the input data frame.

22. The method as defined in claim 21 wherein said computing step comprises selecting a predetermined number of said eigenvectors.

23. A computer system for generating tone-sensitive acoustic models comprising:

a memory unit;

a data input unit to receive input data frames;

a signal processing unit to generate acoustic vectors representing said input data frames; and

a processing unit to,

generate tone-sensitive parameters corresponding to said input data frames,

create feature vectors corresponding to said input data frames by appending said tone-sensitive parameters to said acoustic vectors, and

create acoustic models from said feature vectors.

24. The computer system as defined in claim 23 wherein said signal processing unit is a signal processor.

25. The computer system as defined claim 23 wherein said signal processing unit is a general purpose processor.

26. A computer-readable medium having stored thereon a plurality of instructions for generating tone-sensitive vectors representing frames of input data from which acoustic models can be created, wherein the plurality of instructions are designed to be executed by a processor for implementing a function to create an acoustic vector for a first frame of said input data, to determine a pitch value for said first frame, and to append said pitch value to said acoustic vector for said first frame.

27. The computer-readable medium as defined in claim 26, wherein said instructions to determine said pitch value for said first frame include instructions to use the pitch value detected by a pitch detector for said pitch value for said first frame, provided said pitch value is present in said first frame, and to generate a pitch value for said first frame based on detected pitch values of surrounding frames, provided said pitch is not present for said first frame.