Searching in Audio Speech

• US 20100324900A1
• Filed: 06/19/2009
• Published: 12/23/2010
• Est. Priority Date: 06/19/2009
• Status: Active Grant

• Current Assignee: LNTS Linguistech Solutions Limited
• Original Assignee: LNTS Linguistech Solutions Limited
• Inventors: Simone, Adam, Cohen-Tov, Rabin, Faifkov, Ronen
• Granted Patent: US 8,321,218 B2
• Time in Patent Office: Days
• US Class Current: 704/254
• CPC Class Codes: G10L 15/12 using dynamic programming techniques, e.g. dynamic time warping [DTW]
• CPC Class Codes: G10L 2015/025 Phonemes, fenemes or fenones being the recognition units

Abstract

A computerized method of detecting a target word in a speech signal. A speech recognition engine and a previously constructed phoneme model is provided. The speech signal is input into the speech recognition engine. Based on the phoneme model, the input speech signal is indexed. A time-ordered list is stored representing n-best phoneme candidates of the input speech signal and phonemes of the input speech signal in multiple phoneme frames. The target word is transcribed into a transcription of target phonemes. The time-ordered list of n-best phoneme candidates is searched for a locus of said target phonemes. While searching, scoring is based on the ranking of the phoneme candidates among the n-best phoneme candidates and based on the number of the target phonemes found. A composite score of the probability of an occurrence of the target word is produced. When the composite score is higher than a threshold, start and finish times are output which bound the locus. The start and finish times are input into an algorithm adapted for sequence alignment based on dynamic programming for aligning a portion of the phoneme frames with the target phonemes.

Claims

1. A computerized method of detecting a target word in a speech signal, the method comprising:

providing a speech recognition engine and a previously constructed phoneme model;

inputting the speech signal into the speech recognition engine;

based on the phoneme model, indexing the input speech signal, thereby storing a time-ordered list representing n-best phoneme candidates of the input speech signal and phonemes in a plurality of phoneme frames, wherein n is an integer between two and eight;

transcribing the target word into a transcription of target phonemes;

searching through said time-ordered list of n-best phoneme candidates for a locus of said target phonemes;

while said searching, scoring based on the ranking of the phoneme candidates among the n-best phoneme candidates and based on the number of said target phonemes found, thereby producing a composite score of the probability of an occurrence of the target word;

when said composite score is higher than a threshold, outputting start and finish times bounding said locus;

inputting said start and finish times into an algorithm adapted for sequence alignment based on dynamic programming; and

using said algorithm aligning a first sequence with a second sequence, wherein said first sequence is a portion of said phoneme frames, wherein said portion is based on said start and finish times and wherein said second sequence is said target phonemes.

2. The method of claim 1, wherein said phoneme model explicitly includes as a parameter a length of the phonemes.

3. The method of claim 1, wherein said phoneme model includes dividing a plurality of phonemes into two or three sub-phonemes and modeling said sub-phonemes acoustically using a plurality of Gaussian parameters based on a mixture model of Gaussian functions and an explicit length dependency.

4. The method of claim 3, wherein said length dependency is a Poisson length dependency.

5. The method of claim 1, wherein fewer than fifty phonemes are modeled in said phoneme model.

6. The method of claim 1, wherein said sequence alignment algorithm is based on a Smith-Waterman algorithm.

7. The method of claim 1, wherein said sequence alignment is time constrained to a portion of the speech signal substantially between said start and finish times.

8. The method of claim 1, further comprising, prior to performing said sequence alignment, reducing in said first sequence repetitive frames of the same phoneme to a single frame of the same phoneme.

9. The method of claim 1, wherein storing said phonemes in said phoneme frames includes storing k best candidates of said phonemes in said phoneme frames and wherein said aligning is performed over said k-best candidates, wherein k is an integer between two and eight.

10. The method of claim 1, wherein said threshold is previously determined.

11. A computer readable medium encoded with processing instructions for causing a processor to execute the method of claim 1.