US 8,543,402 B1Granted

Speaker segmentation in noisy conversational speech

AssigneeThe Intellisis Corporation

InventorJiyong Ma

Life of this patent

Estimating the term…

Loading patent…

Abstract

System and methods for robust multiple speaker segmentation in noisy conversational speech are presented. Robust voice activity detection is applied to detect temporal speech events. In order to get robust speech features and detect speech events in a noisy environment, a noise reduction algorithm is applied, using noise tracking. After noise reduction and voice activity detection, the incoming audio/speech is initially labeled as speech segments or silence segments. With no prior knowledge of the number of speakers, the system identifies one reliable speech segment near the beginning of the conversational speech and extracts speech features with a short latency, then learns a statistical model from the selected speech segment. This initial statistical model is used to identify the succeeding speech segments in a conversation. The statistical model is also continuously adapted and expanded with newly identified speech segments that match well to the model. The speech segments with low likelihoods are labeled with a second speaker ID, and a statistical model is learned from them. At the same time, these two trained speaker models are also updated/adapted once a reliable speech segment is identified. If a speech segment does not match well to the two speaker models, the speech segment is temporarily labeled as an outlier or as originating from a third speaker. This procedure is then applied recursively as needed when there are more than two speakers in a conversation.

A method for speaker segmentation and speaker identification of speech input, the method comprising:

(a) segmenting the speech input into a set of silent segments and a set of speech segments;
(b) identifying a first speech segment as corresponding to a first speaker, wherein the first speech segment is included in the set of speech segments;
(c) updating a first speech model corresponding to the first speaker based on the first speech segment, wherein the first speech model is included in a set of speech models;
(d) identifying a second speech segment as corresponding to a second speaker, wherein the second speech segment is included is the set of speech segments, wherein identifying the second speech segment as corresponding to the second speaker is based on a comparison of correlations to individual speech models in the set of speech models;
(e) responsive to (d), updating a second speech model corresponding to the second speaker based on the second speech segment, wherein the second speech model is included in the set of speech models;
(f) correlating an unidentified speech segment from the set of speech segments to one or more speech models in the set of speech models;
(g) responsive to sufficient correlation of the unidentified speech segment with an individual speech model:
(1) identify the unidentified speech segment as corresponding to an individual speaker, wherein the individual speaker corresponds to the individual speech model,
(2) updating the individual speech model based on the unidentified speech segment, and
(3) marking the unidentified speech segment as an identified speech segment;
(h) responsive to insufficient correlation, subsequent to step (f), of the unidentified speech segment with an individual speech model:
(1) adding a new individual speech model to the set of speech models,
(2) updating the new individual speech model based on the unidentified speech segment, and
(3) marking the unidentified speech segment as an identified speech segment;
(i) repeating steps (f), (g), and (h) until individual ones of the speech segments in the set of speech segments are identified as corresponding to individual speakers that correspond to individual ones of the speech models in the set of speech models;
and
(j) for a particular speech segment in the set of speech segments that has previously been identified as corresponding to a particular speech model, identifying a specific speech model from the set of speech models that is different than the particular speech model, wherein the specific speech model has a higher correlation to the particular speech segment than the particular speech model.

14 more elements

Facts

Application

13/097,620Filed 29 Apr 2011 · us provisional application

Priority

30 Apr 2010

Granted

24 Sep 2013

Term ends

Not estimatedThe source records don’t support an estimate.

Classification3 codes

G10L 17/04
Speaker identification or verification techniques: training, enrolment or model building
G10L 17/08
Decision making techniques; Pattern matching strategies: use of distortion metrics or a particular distance between probe pattern and reference…
G10L
Speech analysis techniques or speech synthesis

Abstract

Claims6