Unsupervised labeling of sentence level accent

• US 20080201145A1
• Filed: 02/20/2007
• Published: 08/21/2008
• Est. Priority Date: 02/20/2007
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Chu, Min, Soong, Frank Kao-ping, Chen, YiNing
• Granted Patent: US 7,844,457 B2
• Time in Patent Office: Days
• US Class Current: 704/244
• CPC Class Codes: G10L 13/08 Text analysis or generation of parameters for speech synthesis out of text, e.g. grapheme to phoneme translation, prosody generation or stress or intonation determination

Abstract

Methods are disclosed for automatic accent labeling without manually labeled data. The methods are designed to exploit accent distribution between function and content words.

Claims

1. A method of training an acoustic model, the method comprising:

classifying each of a plurality of words as being either a content word or a function word;

utilizing a characteristic of at least one of the content words as a basis for identifying an accent characteristic of at least one of the function words; and

training the acoustic model so as to be indicative of the accent characteristic.

2. The method of claim 1, wherein training the acoustic model comprises training so as to add an indication of an unaccented vowel of a word that has been classified as a function word.

3. The method of claim 2, further comprising training the acoustic model so as to add an indication of an accented vowel of a word that has been classified as a content word.

4. The method of claim 1, further comprising utilizing the characteristic as a basis for labeling an accent characteristic of at least one of the function words.

5. The method of claim 1, wherein utilizing a characteristic of at least one of the content words as a basis for identifying an accent characteristic of at least one of the function words comprises utilizing accented and unaccented vowels.

6. The method of claim 1, wherein training the acoustic model comprises training the acoustic model to be indicative of accented vowels of words that have been classified as being content words, and also indicative of unaccented vowels of words that have been classified as being function words.

7. The method of claim 6, wherein accented vowels of words that have been labeled as function words are excluded from the collection of data utilized to train the acoustic model.

8. The method of claim 1, further comprising utilizing the acoustic model as a basis for labeling (208) the function word.

9. A method of training an acoustic model, the method comprising:

utilizing a first acoustic model to label accented and unaccented components of function words; and

utilizing the unaccented components as a basis for training a second acoustic model.

10. The method of claim 9, further comprising utilizing accented components of content words as a basis for training the second acoustic model.

11. The method of claim 9, further comprising excluding the accented components of the function words from the collective set of data utilized as a basis for training the second acoustic model.

12. The method of claim 9, wherein the first acoustic model contains a representation of accented and unaccented components of words that have been identified as being content words.

13. The method of claim 9, further comprising utilizing the second acoustic model as a basis for labeling accented and unaccented components of words that have been identified as being function words.

14. A method of generating an acoustic model, the method comprising utilizing accented vowels that are part of words identified as being content words, as well as unaccented vowels that are part of words identified as being functions words, as a basis for generating the acoustic model.

15. The method of claim 14, further comprising utilizing the acoustic model as a basis for labeling a new set of accented and unaccented components of words that have been identified as being function words.

16. The method of claim 15, further comprising utilizing the unaccented components from the new set as a basis for generating a refined acoustic model.

17. The method of claim 14, wherein generating comprises training a hidden-markov-model.

18. The method of claim 14, wherein utilizing accented vowels comprises utilizing vowels identified as accented through application of a model trained based on characteristics of a collection of words identified as being content words.

19. The method of claim 18, wherein utilizing vowels identified as accented through application of a model trained based on characteristics of words identified as being content words comprises utilizing vowels identified as accented through application of a model trained to distinguish between accented and unaccented vowels of a collection of words identified as being content words.

20. The method of claim 14, wherein the method is computer-implemented without manual intervention.