Back to results
US 12,718,821 B2Granted
Method and apparatus for speaker verification based on next-TDNN
AssigneesHyundai Motor Company, Kia Corporation, Sogang University Research & Business Development Foundation
InventorsRan Lee, Hyung-Min Park, Hyun-Jun Heo, Ui-Hyeop Shin
Life of this patent
Estimating the term…
Abstract
A method and an apparatus are disclosed for a speaker verification. The apparatus may comprise at least one processor configured to execute instructions to perform generating, based on an utterance of a speaker to a deep learning-based speaker verification model, a speaker embedding with a preset dimension, verifying, based on the speaker embedding, the speaker, and detecting, based on the verified speaker, an identity of a user associated with the apparatus.
Claims20
An apparatus comprising:
- at least one processor;
- and
- a memory that stores instructions and communicates with the at least one processor, wherein the at least one processor is configured to execute the instructions to perform:
- generating, based on an utterance of a speaker to a deep learning-based speaker verification model, a speaker embedding with a preset dimension by:
- generating first output frames by providing a plurality of input frames to one dimensional convolution layer, wherein the first output frames comprise features of a predefined dimension;
- providing the first output frames to two-step layers, wherein each of the two-step layers is configured to generate second output frames comprising multi-scale features, wherein the two-step layers are connected in series, wherein each of the two-step layers comprises at least one two-step block, and wherein the at least one two-step block is connected in series;
- generating third output frames comprising aggregated features by concatenating an output of each of the two-step layers and by providing the concatenated output to a multi-layer feature aggregation block;
- generating an attention vector by providing the third output frames to an attentive statistics pooling (ASP) block;
- weighted-summing, based on the attention vector, the third output frames;
- and
- generating the speaker embedding by providing an output of the ASP block to a linear layer, wherein the output of the ASP block is generated based on the weighted-summing;
- verifying, based on the speaker embedding, the speaker;
- and
- detecting, based on the verified speaker, an identity of a user associated with the apparatus.
11 more elements
A method performed by a processor, the method comprising:
- generating, based on an utterance of a speaker to a deep learning-based speaker verification model, a speaker embedding with a preset dimension by:
- generating first output frames by providing a plurality of input frames to one dimensional convolution layer, wherein the first output frames comprises features of a predefined dimension;
- providing the first output frames to two-step layers, wherein each of the two-step layers generates second output frames comprising multi-scale features, wherein the two-step layers are connected in series, wherein each of the two-step layers comprises at least one two-step block, and wherein the at least one two-step block is connected in series;
- generating third output frames comprising aggregated features by concatenating an output of each of the two-step layers and by providing the concatenated output to a multi-layer feature aggregation block;
- generating an attention vector by providing the third output frames to an attentive statistics pooling (ASP) block;
- weighted-summing, based on the attention vector, the third output frames;
- and
- generating the speaker embedding by providing an output of the ASP block to a linear layer, wherein the output of the ASP block is generated based on the weighted-summing;
- verifying, based on the speaker embedding, the speaker;
- and
- detecting, based on the verified speaker, an identity of a user associated with the apparatus.
8 more elements
A non-transitory computer-readable recording medium storing instructions, when executed, cause:
- generating, based on an utterance of a speaker to a deep learning-based speaker verification model, a speaker embedding with a preset dimension by:
- generating first output frames by providing a plurality of input frames to one dimensional convolution layer, wherein the first output frames comprise features of a predefined dimension;
- providing the first output frames to two-step layers, wherein each of the two-step layers configured to generate second output frames comprising multi-scale features, wherein the two-step layers are connected in series, wherein each of the two-step layers comprises at least one two-step block, and wherein the at least one two-step block is connected in series;
- generating third output frames comprising aggregated features by concatenating an output of each of the two-step layers and by providing the concatenated output to a multi-layer feature aggregation block;
- generating an attention vector by providing the third output frames to an attentive statistics pooling (ASP) block;
- weighted-summing, based on the attention vector, the third output frames;
- and
- generating the speaker embedding by providing an output of the ASP block to a linear layer, wherein the output of the ASP block is generated based on the weighted-summing;
- verifying, based on the speaker embedding, the speaker;
- and
- detecting, based on the verified speaker, an identity of a user associated with the apparatus.
8 more elements
Dependencies are inferred from the claim text.