ACCURACY IMPROVEMENT OF SPOKEN QUERIES TRANSCRIPTION USING CO-OCCURRENCE INFORMATION

• US 20140136197A1
• Filed: 01/16/2014
• Published: 05/15/2014
• Est. Priority Date: 07/31/2011
• Status: Active Grant

• Current Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Original Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Inventors: Mamou, Jonathan, Sethy, Abhinav, Ramabhadran, Bhuvana, Hoory, Ron, Vozila, Paul Joseph, Bodenstab, Nathan
• Granted Patent: US 9,330,661 B2
• Time in Patent Office: Days
• US Class Current: 704/235
• CPC Class Codes: G06F 16/00 Information retrieval; Database structures therefor; File system structures therefor
• CPC Class Codes: G06F 7/00 Methods or arrangements for processing data by operating upon the order or content of the data handled logic circuits H03K19/00
• CPC Class Codes: G10L 15/08 Speech classification or search
• CPC Class Codes: G10L 15/1815 Semantic context, e.g. disambiguation of the recognition hypotheses based on word meaning
• CPC Class Codes: G10L 15/26 Speech to text systems G10L15/08 takes precedence

Abstract

Techniques disclosed herein include systems and methods for voice-enabled searching. Techniques include a co-occurrence based approach to improve accuracy of the 1-best hypothesis for non-phrase voice queries, as well as for phrased voice queries. A co-occurrence model is used in addition to a statistical natural language model and acoustic model to recognize spoken queries, such as spoken queries for searching a search engine. Given an utterance and an associated list of automated speech recognition n-best hypotheses, the system rescores the different hypotheses using co-occurrence information. For each hypothesis, the system estimates a frequency of co-occurrence within web documents. Combined scores from a speech recognizer and a co-occurrence engine can be combined to select a best hypothesis with a lower word error rate.

Claims

1. A computer-implemented method for executing a voice search, the computer-implemented method comprising:

receiving a spoken query;

identifying, via an automated speech recognition process, multiple transcription hypotheses based on the spoken query, each respective transcription hypothesis having a speech recognition score;

evaluating a plurality of the transcription hypotheses using a co-occurrence identification process, the co-occurrence identification process;

identifying a frequency that proposed query terms, from each respective transcription hypothesis, co-occur based on a corpus of documents;

assigning a co-occurrence score to each respective transcription hypothesis; and

selecting a best transcription hypothesis based on at least non-sequential co-occurrences of the proposed query terms within the corpus documents; and

generating a text query corresponding to the best transcription hypothesis.

2. The computer-implemented method of claim 1 wherein the co-occurrence score represents a measure of semantic relation of the proposed query terms based on identified co-occurrence frequencies including identifying non-sequential co-occurrences of the proposed query terms within the documents in the corpus.

3. The computer-implemented method of claim 2 further comprising selecting a best transcription hypothesis based on a combination of speech recognition scores and co-occurrence scores from the plurality of the transcription hypotheses;

and

receiving search results, from an information retrieval system, based on a text query corresponding to the best transcription hypothesis.

4. The computer-implemented method of claim 1, wherein identifying the frequency that proposed query terms co-occur within documents in the corpus includes identifying co-occurrences of the proposed query terms within a predetermined number of consecutive words in the individual corpus documents.

5. The computer-implemented method of claim 1, wherein evaluating the plurality of the transcription hypotheses using the co-occurrence identification process includes the plurality being a group of transcription hypotheses selected as having best speech recognition scores based on a predetermined criterion.

6. The computer-implemented method of claim 1, wherein selecting the best transcription hypothesis is based on a combination of speech recognition scores and co-occurrence scores from the plurality of the transcription hypotheses, further including:

evaluating a weighting of the speech recognition scores and the co-occurrence scores; and

in response to identifying that a given hypothesis has a co-occurrence score based on proposed query terms identified as co-occurring within a predetermined number of consecutive words within individual corpus documents, increasing a weighting of the co-occurrence score relative to a baseline weighting.

7. The computer-implemented method of claim 6, wherein selecting the best transcription hypothesis based on a combination of speech recognition scores and co-occurrence scores from the plurality of the transcription hypotheses includes:

rescoring respective speech recognition scores based on corresponding co-occurrence scores; and

identifying the best transcription hypothesis as having a highest score based on the rescoring.

8. The computer-implemented method of claim 1, wherein selecting the best transcription hypothesis is based on the combination of speech recognition scores and co-occurrence scores from the plurality of the transcription hypotheses, further including:

evaluating a weighting of the speech recognition scores and the co-occurrence scores; and

in response to identifying that a given hypothesis has over a predetermined number of proposed query terms, weighting a co-occurrence score of the given hypothesis less than a weight of the co-occurrence score when the given hypothesis has less then the-predetermined number of proposed query terms.

9. The computer-implemented method of claim 1, wherein identifying the multiple transcription hypotheses via the automated speech recognition process includes the automated speech recognition process analyzing a waveform of the spoken query using an acoustic language model and a sequence-based statistical language model.

10. The computer-implemented method of claim 9, wherein the statistical language model is trained on a first text corpus of natural language utterances and a second text corpus of search engine queries.

11. The computer-implemented method of claim 1, wherein the spoken query is received from a voice search interface of a mobile client device.

12. The computer-implemented method of claim 11, further comprising:

displaying the search results via the mobile client device.

13. The computer-implemented method of claim 1, further comprising:

evaluating the plurality of the transcription hypotheses using a class identification process, the class identification process including determining that a given query term, from a respective transcription hypothesis, corresponds to a specific class of terms, the class identification process assigning a classification score to the given query term; and

wherein selecting the best transcription hypothesis includes basing selection on classification scores.

14. The computer-implemented method of claim 1, further comprising:

evaluating the plurality of the transcription hypotheses using a word relatedness identification process, the word relatedness identification process including evaluating a given query term, from a respective transcription hypothesis, using a lexical database, the word relatedness identification process assigning a word relatedness score to the given query term, the word relatedness score indicating a measure of semantic relation of the given query term to other words; and

wherein selecting the best transcription hypothesis includes basing selection on word relatedness scores.

15. A computer system for executing a voice search, the computer system comprising:

a processor; and

a memory coupled to the processor, the memory storing instructions that, when executed by the processor, cause the system to perform the operations of;

receiving a spoken query;

identifying, via an automated speech recognition process, multiple transcription hypotheses based on the spoken query, each respective transcription hypothesis having a speech recognition score;

evaluating a plurality of the transcription hypotheses using a co-occurrence identification process, the co-occurrence identification process;

identifying a frequency that proposed query terms, from each respective transcription hypothesis, co-occur based on a corpus of documents;

assigning a co-occurrence score to each respective transcription hypothesis; and

selecting a best transcription hypothesis based on at least non-sequential co-occurrences of the proposed query terms within the corpus documents; and

generating a text query corresponding to the best transcription hypothesis.