Back to results
US 10,832,685 B2Granted
Speech processing device, speech processing method, and computer program product
AssigneeKabushiki Kaisha Toshiba
InventorsNing Ding, Makoto Hirohata
Life of this patent
Estimating the term…
Abstract
According to an embodiment, a speech processing device includes an extractor, a classifier, a similarity calculator, and an identifier. The extractor is configured to extract a speech feature from utterance data. The classifier is configured to classify the utterance data into a set of utterances for each speaker based on the extracted speech feature. The similarity calculator is configured to calculate a similarity between the speech feature of the utterance data included in the set and each of a plurality of speaker models. The identifier is configured to identify a speaker for each set based on the calculated similarity.
Claims9
A speech processing device comprising:
- a memory:
- a processor electrically coupled to the memory, and configured to function as an extractor, a classifier, a similarity calculator, and an identifier, wherein
- the extractor extracts a speech feature from utterance data;
- the classifier classifies the utterance data into a set of utterances for each speaker based on the extracted speech feature;
- the similarity calculator calculates a similarity between the speech feature of the utterance data included in the set and each of a plurality of speaker models;
- and
- the identifier identifies a speaker for each set based on the calculated similarity, wherein
- the identifier (1) identifies the speaker in the set having a weighted sum of the similarities equal to or larger than a first threshold as a known speaker corresponding to one of the speaker models, and (2) the identifier identifies the speaker in the set having a weighted sum of the similarities smaller than a second threshold, the second threshold being less than the first threshold, as a new speaker, and
- the processor is configured to further function as a generator, and the generator generates a new speaker model in the set having the weighted sum of the similarities that is smaller than the second threshold for the identified new speaker,
- wherein the classifier classifies the utterance data into clusters based on similarity of the utterance data, and the identifier obtains a weighted sum of similarity between each cluster and each speaker model, and when the weighted sum exceeds the first threshold, the respective cluster is identified as the known speaker for a corresponding speaker model, and when multiple clusters are identified as a same known speaker, the multiple clusters are integrated for the same known speaker.
7 more elements
A speech processing method comprising:
- extracting a speech feature from utterance data;
- classifying the utterance data into a set of utterances for each speaker based on the extracted speech feature;
- calculating a similarity between the speech feature of the utterance data included in the set and each of a plurality of speaker models;
- and
- identifying a speaker for each set based on the calculated similarity, wherein
- the identifying includes (1) identifying the speaker in the set having a weighted sum of the similarities equal to or larger than a first threshold as a known speaker corresponding to one of the speaker models, and (2) the identifier identifies the speaker in the set having a weighted sum of the similarities smaller than a second threshold, the second threshold being less than the first threshold, as a new speaker, and
- the speech processing method further comprises generating a new speaker model in the set having the weighted sum of the similarities that is smaller than the second threshold for the identified new speaker,
- wherein the classifying classifies the utterance data into clusters based on similarity of the utterance data, and the identifying obtains a weighted sum of similarity between each cluster and each speaker model, and when the weighted sum exceeds the first threshold, the respective cluster is identified as the known speaker for a corresponding speaker model, and when multiple clusters are identified as a same known speaker, the multiple clusters are integrated for the same known speaker.
5 more elements
A computer program product comprising a non-transitory computer-readable medium containing a program, the program causing a computer to execute:
- extracting a speech feature from utterance data;
- classifying the utterance data into a set of utterances for each speaker based on the extracted speech feature;
- calculating a similarity between the speech feature of the utterance data included in the set and each of a plurality of speaker models;
- and
- identifying a speaker for each set based on the calculated similarity, wherein
- the identifying includes (1) identifying the speaker in the set having a weighted sum of the similarities equal to or larger than a first threshold as a known speaker corresponding to one of the speaker models, and (2) the identifier identifies the speaker in the set having a weighted sum of the similarities smaller than a second threshold, the second threshold being less than the first threshold, as a new speaker, and
- the program causes the computer to further execute:
- generating a new speaker model in the set having the weighted sum of the similarities that is smaller than the second threshold for the identified new speaker,
- wherein the classifier classifying the utterance data into clusters based on similarity of the utterance data, and the identifying obtains a weighted sum of similarity between each cluster and each speaker model, and when the weighted sum exceeds the first threshold, the respective cluster is identified as the known speaker for a corresponding speaker model, and when multiple clusters are identified as a same known speaker, the multiple clusters are integrated for the same known speaker.
6 more elements
Dependencies are inferred from the claim text.