Method and system for creating frugal speech corpus using internet resources and conventional speech corpus

• US 8,756,064 B2
• Filed: 06/26/2012
• Issued: 06/17/2014
• Est. Priority Date: 07/28/2011
• Status: Active Grant

• Current Assignee: TATA Consultancy Services Limited (Tata Sons Pvt Ltd.)
• Original Assignee: TATA Consultancy Services Limited (Tata Sons Pvt Ltd.)
• Inventors: Sheikh, Imran Ahmed, Kopparapu, SunilKumar
• Primary Examiner(s): Opsasnick, Michael N
• Application Number:
  US13/533,174
• Publication Number:
  US 20130030810A1
• Time in Patent Office: 721 Days
• Field of Search: 704/260
• US Class Current: 704/260
• CPC Class Codes: G06F 16/954 Navigation, e.g. using categorised browsing
• CPC Class Codes: G10L 15/06 Creation of reference templates; Training of speech recognition systems, e.g. adaptation to the characteristics of the speaker's voice G10L15/14 takes precedence

Abstract

A speech corpus creation method and system are disclosed. The method comprising identifying a publicly accessible first source of the first speech data and its corresponding first text transcription; extracting a second speech data of an accessible encoding format from the first speech data; extracting a second text transcription data with at least one encoding format from the first text transcription data; matching and aligning the transcription to the extracted second speech data at a sentence, word, phoneme level, or combination thereof to form a first and a second speech corpus; analyzing the text transcriptions in the second speech corpus to identify the short speech segments to produce a phonetically balanced, segmented, text aligned third speech corpus; and conditioning the third speech corpus by inserting a context and associated environment richer corpus therein the third speech corpus from at least one second source to form the final speech corpus.

Claims

1. A speech corpus creation method, implementing extraction of a first speech data from at least one first source and mixing with at least one second source, the method comprising processor implemented steps of:

identifying at least one publicly accessible first source of the first speech data and its corresponding first text transcription;

extracting a second speech data of at least one accessible encoding format from the first speech data;

extracting a second text transcription data with at least one encoding format from the first text transcription data;

matching and aligning the transcription to the extracted second speech data at a sentence, word, phoneme level, or combination thereof to form a first and a second speech corpus;

analyzing the text transcriptions in the second speech corpus to identify the short speech segments to produce a phonetically balanced, segmented, text aligned third speech corpus; and

conditioning the third speech corpus by inserting a context and associated environment richer corpus therein the third speech corpus from at least one second source to form the final speech corpus.

2. A method as claimed in claim 1, wherein speech data extractor engine extracts the first speech data along with transcription thereof from publicly accessible sources that are relevant to a desired corpus.

3. A method as claimed in claim 1, wherein while aligning the speech, a two stage alignment process is carried out comprising a first level syllable matching adapted to match the extracted transcription to the corresponding extracted long speech data at a syllable level and a subsequent second level matching adapted to align, using automatic speech recognition engine, short speech segments of the long syllable aligned speech data at sentence, word or phoneme level.

4. A method as claimed in claim 1, wherein the matching and aligning comprising steps of:

detecting plurality of syllables in the second speech data;

detecting plurality of syllables in the second text transcription data by employing a text syllable annotator;

annotating and indexing each detected syllable in the second speech data and in the second text transcription data;

aligning the syllable annotated second speech data with the syllable annotated second text data by matching the corresponding syllable indexes, to form a first syllable aligned speech corpus;

segmenting the said first aligned corpus into plurality of short speech segments of uniform length; and

aligning each short segment with the corresponding exacted text transcription to form a segmented text aligned second speech corpus, featuring alignment at sentence, word or phoneme level.

5. A method as claimed in claim 1, wherein the third speech corpus, derived from a public source of speech data and its transcription, is conditioned with a context and associated environment richer corpus, collected using traditional procedure, to form the final speech corpus.

6. A speech corpus creation system, implementing extraction of a first speech data from at least one first source and mixing with at least one second source, the system comprising:

a speech data extractor adapted to extract a second speech data of at least one encoding format from the first speech data;

a text data extractor adapted to extract a second text transcription data of at least one encoding format from a first text transcription of the first speech data;

a speech alignment module adapted to match and align the first text transcription to the corresponding extracted long speech data in the first speech data, at a sentence word level, or combination thereof to form a first and a second speech corpus;

a phonetically balanced data extractor for analyzing the text transcriptions in the second speech corpus and to identify the short speech segments to form a phonetically balanced, segmented, text aligned third speech corpus; and

a compensator means adapted to identify at least one contextual gap in the third speech corpus and to condition the third speech corpus by inserting a context and associated environment richer corpus therein the third speech corpus from the at least one second source to form a final speech corpus.

7. A system as claimed in claim 6, wherein speech data extractor engine extracts the first speech data along with transcription thereof from publicly accessible sources that are relevant to a desired corpus domain.

8. A system as claimed in claim 6, wherein while aligning the speech, a two stage alignment is carried out comprising a first level syllable matching adapted to match the extracted transcription to the corresponding extracted long speech data at a syllable level and a subsequent second level of matching adapted to align, using automatic speech recognition, short speech segments of the syllable aligned long speech data, at least sentence, word and phoneme level.

9. A system as claimed in claim 6, wherein the speech alignment module comprising of:

a speech syllable annotator adapted to annotate and index plurality of syllables in the second speech data;

a text syllable annotator adapted to annotate and index the syllables in the second text transcription data;

a syllable based aligner adapted to align the syllable indexed second speech data to the syllable indexed second text data by matching syllable indexes, to form a first syllable aligned long speech corpus, a long speech Segmenter adapted to segment the first syllable aligned long speech corpus into plurality of uniform segments; and

a short speech aligner adapted to align each short speech segment at least sentence, word and phoneme level with the corresponding transcription using an automatic speech recognition engine to form a segmented text aligned second speech corpus.