Diarization driven by meta-information identified in discussion content

Patent metadata

Field	Value
Publication number	US-10468031-B2
Application number	US-201715819158-A
Country	US
Kind code	B2
Filing date	Nov 21, 2017
Priority date	Nov 21, 2017
Publication date	Nov 5, 2019
Grant date	Nov 5, 2019

Abstract

An approach is provided that receives an audio stream and utilizes a voice activation detection (VAD) process to create a digital audio stream of voices from at least two different speakers. An automatic speech recognition (ASR) process is applied to the digital stream with the ASR process resulting in the spoken words to which a speaker turn detection (STD) process is applied to identify a number of speaker segments with each speaker segment ending at a word boundary. The STD process analyzes a number of speaker segments using a language model that determines when speaker changes occur. A speaker clustering algorithm is then applied to the speaker segments to associate one of the speakers with each of the speaker segments.

First claim

The invention claimed is:

1. A method implemented by an information handling system that includes a memory and a processor, the method comprising:
receiving an audio stream that comprises both a plurality of speech segments corresponding to a plurality of human speakers and a plurality of non-verbal segments;
utilizing a voice activation detection (VAD) process on the audio stream, wherein an output of the VAD process is a digital audio stream of voices corresponding to the plurality of speech segments;
applying an automatic speech recognition (ASR) process to the digital stream, wherein the ASR process results in a plurality of spoken words;
inputting the VAD process output into an automatic speech recognition (ASR) process, wherein an output of the ASR process comprises a plurality of spoken words corresponding to the plurality of speech segments and is devoid of the plurality of non-verbal segments
inputting the ASR process output to a speaker turn detection (STD) process, wherein a plurality of speaker segments of contiguous words are selected from the plurality of spoken words and analyzed by a language model that determines when a plurality of speaker changes occur based on meta-information corresponding to the plurality of speaker segments, the analyzing further comprising:
associating a first word from the plurality of spoken words to a first set of vocal qualities;
identifying a second word from the plurality of spoken words that is successive to the first word and corresponds to a second set of vocal qualities;
inserting a speaker change mark between the first word and the second word in response to determining that the first set of vocal qualities is different from the second set of vocal qualities;
increasing a speaker change value in response to determining that a selected one of the speaker segments corresponding to the first word is a question; and
confirming the speaker change mark in response to determining that a speaker change occurs based on the increased speaker change value; and

in response to confirming the speaker change mark, applying a speaker clustering algorithm to the plurality of speaker segments, wherein the speaker clustering algorithm associates an identifier of one of the human speakers with each of the speaker segments.

2. The method of claim 1 further comprising:
increasing the speaker change value in response to the language model analysis revealing that the selected speaker segment is a statement;
increasing the speaker change value in response to the language model analysis revealing that the selected speaker segment is a reply; and
decreasing the speaker change value in response to the language model analysis revealing that the selected speaker segment is a continuation of one or more of the previous speaker segments.

3. The method of claim 1 further comprising:
appending the selected speaker segment to the set of previous speaker segments;
selecting a next one of the speaker segments;
analyzing the selected next speaker segment during the STD process based on the set of previous speaker segment that now includes the selected speaker segment;
analyzing a second speaker change value based on the language model analysis; and
determining whether a second speaker change occurs based on the second speaker change value.

4. The method of claim 1 wherein a plurality of sets of vocal qualities comprise the first set of vocal qualities and the second set of vocal qualities, the method further comprising:
identifying the plurality of sets of vocal qualities from the audio stream, wherein each of the sets of vocal qualities corresponds to a different one of the plurality of human speakers;
comparing the plurality of sets of vocal qualities to each of the plurality of spoken words; and
associating one of the human speakers to each of the words based on the comparison.

5. The method of claim 1 further comprising:
generating a transcript of the audio stream that includes the plurality of speaker segments and an association of each of the speaker segments to one of the human speakers; and
ingesting the transcript into a corpus utilized by a question answering (QA) system.

6. An information handling system comprising:
one or more processors;
a memory coupled to at least one of the processors; and
a set of computer program instructions stored in the memory and executed by at least one of the processors in order to perform actions of:
receiving an audio stream that comprises both a plurality of speech segments corresponding to a plurality of human speakers and a plurality of non-verbal segments;
utilizing a voice activation detection (VAD) process on the audio stream, wherein an output of the VAD process is a digital audio stream of voices corresponding to the plurality of speech segments;
applying an automatic speech recognition (ASR) process to the digital stream, wherein the ASR process results in a plurality of spoken words;
inputting the VAD process output into an automatic speech recognition (ASR) process, wherein an output of the ASR process comprises a plurality of spoken words corresponding to the plurality of speech segments and is devoid of the plurality of non-verbal segments
inputting the ASR process output to a speaker turn detection (STD) process, wherein a plurality of speaker segments of contiguous words are selected from the plurality of spoken words and analyzed by a language model that determines when a plurality of speaker changes occur based on meta-information corresponding to the plurality of speaker segments, the analyzing further comprising:
associating a first word from the plurality of spoken words to a first set of vocal qualities;
identifying a second word from the plurality of spoken words that is successive to the first word and corresponds to a second set of vocal qualities;
inserting a speaker change mark between the first word and the second word in response to determining that the first set of vocal qualities is different from the second set of vocal qualities;
increasing a speaker change value in response to determining that a selected one of the speaker segments corresponding to the first word is a question; and
confirming the speaker change mark in response to determining that a speaker change occurs based on the increased speaker change value; and

in response to confirming the speaker change mark, applying a speaker clustering algorithm to the plurality of speaker segments, wherein the speaker clustering algorithm associates an identifier of one of the human speakers with each of the speaker segments.

7. The information handling system of claim 6 wherein the actions further comprise:
increasing the speaker change value in response to the language model analysis revealing that the selected speaker segment is a statement;
increasing the speaker change value in response to the language model analysis revealing that the selected speaker segment is a reply; and
decreasing the speaker change value in response to the language model analysis revealing that the selected speaker segment is a continuation of one or more of the previous speaker segments.

8. The information handling system of claim 6 wherein the actions further comprise:
appending the selected speaker segment to the set of previous speaker segments;
selecting a next one of the speaker segments;
analyzing the selected next speaker segment during the STD process based on the set of previous speaker segment that now includes the selected speaker segment;
analyzing a second speaker change value based on the language model analysis; and
determining whether a second speaker change occurs based on the second speaker change value.

9. The information handling

Assignees

• IBM

Inventors

• Church Kenneth W
• Dimitriadis Dimitrios B
• Fousek Petr
• Novak Miroslav
• Saon George A

Classifications

• G10L17/04

  Training, enrolment or model building · CPC title

• G06F40/35

  Discourse or dialogue representation · CPC title

• G10L17/06

  Decision making techniques; Pattern matching strategies · CPC title

• G10L15/26 Primary

  Speech to text systems (G10L15/08 takes precedence) · CPC title

• G10L15/30

  Distributed recognition, e.g. in client-server systems, for mobile phones or network applications · CPC title