Cross-lingual speaker recognition
Disclosed are systems and methods including computing-processes executing machine-learning architectures for voice biometrics, in which the machine-learning architecture implements one or more language compensation functions. Embodiments include an embedding extraction engine (sometimes referred to as an “embedding extractor”) that extracts speaker embeddings and determines a speaker similarity score for determine or verifying the likelihood that speakers in different audio signals are the same speaker. The machine-learning architecture further includes a multi-class language classifier that determines a language likelihood score that indicates the likelihood that a particular audio signal includes a spoken language. The features and functions of the machine-learning architecture described herein may implement the various language compensation techniques to provide more accurate speaker recognition results, regardless of the language spoken by the speaker.
1. A computer-implemented method comprising: extracting, by a computer, an enrolled voiceprint for an enrolled speaker by applying an embedding extraction engine on one or more enrollment signals of the enrolled speaker, the enrolled voiceprint representing a plurality of enrollment acoustic features of the one or more enrollment signals; extracting, by the computer, an inbound voiceprint for an inbound speaker by applying the embedding extraction engine on one or more inbound signals of the inbound speaker, the inbound voiceprint representing a plurality of inbound acoustic features of the one or more inbound signals; generating, by the computer, one or more language likelihood scores by applying a language classifier on the enrolled voiceprint and the inbound voiceprint indicating a likelihood that an enrollment signal and a paired inbound signal include one or more languages; and generating, by the computer, a cross-lingual quality measure based upon one or more differences of the one or more language likelihood scores generated for the one or more enrollment signals and the one or more inbound signals, the cross-lingual quality measure indicating whether the enrollment signal and the paired inbound signal include a same language of the one or more languages.
2. The method according to claim 1 , further comprising: generating, by the computer, a speaker verification score for the inbound speaker based upon a distance between the enrolled voiceprint and the inbound voiceprint; and updating, by the computer, the speaker verification score according to the cross-lingual quality measure for calibrating the distance.
3. The method according to claim 2 , further comprising identifying, by the computer, the inbound speaker as the enrolled speaker in response to determining that the speaker verification score satisfies a verification threshold.
4. The method according to claim 1 , further comprising: obtaining, by the computer, a plurality of training audio signals having one or more utterances of the one or more languages and a plurality of training labels corresponding to the plurality of training audio signals, each training label indicating the one or more languages in a corresponding audio training audio signal; extracting, by the computer, a plurality of training embeddings by applying the embedding extraction engine on the plurality of training audio signals; and training, by the computer, the language classifier for determining a language likelihood score by applying the language classifier on the plurality of training embeddings.
5. The method according to claim 1 , further comprising: obtaining, by the computer, a plurality of training audio signals having one or more utterances of the one or more languages and a plurality of training labels corresponding to the plurality of training audio signals, each training label indicating the one or more languages in a corresponding training audio signal; extracting, by the computer, a plurality of training embeddings by applying the embedding extraction engine on the plurality of training audio signals; and training, by the computer, one or more layers of the embedding extraction engine by applying a loss function on the plurality of training embeddings and the plurality of training labels corresponding to the plurality of training audio signals.
6. The method according to claim 5 , further comprising retraining, by the computer, the one or more layers of the embedding extraction engine by applying the loss function on the cross-lingual quality measure, the enrolled voiceprint, and a speaker verification score.
7. The method according to claim 5 , further comprising: selecting, by the computer, a subset of training audio signals having a plurality of languages of the plurality of training audio signals according to a subset of training labels of the plurality of training labels; and retraining, by the computer, the one or more layers of the embedding extraction engine by applying the loss function on a subset of training embeddings corresponding to the subset of training audio signals and the subset of training labels.
8. The method according to claim 5 , further comprising, for each training audio signal of the plurality of training audio signals: extracting, by the computer, a plurality of features from the training audio signal; and generating, by the computer, an augmentation training signal corresponding to the training audio signal, the augmentation training signal comprising a plurality of reversed features corresponding to the plurality of features of the corresponding training audio signal.
9. A system comprising: a non-transitory storage configured to store machine-executable instructions of an embedding extractor and a language classifier; and a computer in communication with the non-transitory storage, comprising a processor and configured to: extract an enrolled voiceprint for an enrolled speaker by applying the embedding extractor on one or more enrollment signals of the enrolled speaker, the enrolled voiceprint representing a plurality of enrollment acoustic features of the one or more enrollment signals; extract an inbound voiceprint for an inbound speaker by applying the embedding extractor extraction engine on one or more inbound signals of the inbound speaker, the inbound voiceprint representing a plurality of inbound acoustic features of the one or more inbound signals; generate one or more language likelihood scores by applying the language classifier on the enrolled voiceprint and the inbound voiceprint indicating a likelihood that an enrollment signal and the inbound voiceprint include one or more languages; generate a cross-lingual quality measure based upon one or more differences of the one or more language likelihood scores generated for the one or more enrollment signals and the one or more inbound signals, the cross-lingual quality measure indicating a likelihood that the enrollment signal and a paired inbound signal include a same language of the one or more languages; generate a speaker verification score for the inbound speaker based upon a distance between the enrolled voiceprint and the inbound voiceprint; and update the speaker verification score according to the cross-lingual quality measure for calibrating the distance.
10. The system according to claim 9 , wherein the computer is further configured to: generate a speaker verification score for the inbound speaker based upon a distance between the enrolled voiceprint and the inbound voiceprint; update the speaker verification score according to the cross-lingual quality measure for calibrating the distance; and identify the inbound speaker as the enrolled speaker in response to determining that the speaker verification score satisfies a verification threshold.
11. The system according to claim 9 , wherein the computer is further configured to: obtain a plurality of training audio signals having one or more utterances of the one or more languages and a plurality of training labels corresponding to the plurality of training audio signals, each training label indicating the one or more languages in a corresponding audio training audio signal; extract a plurality of training embeddings by applying the embedding extractor extraction engine on the plurality of training audio signals; and train the language classifier for determining a language likelihood score by applying the language classifier on the plurality of training embeddings.
12. The system according to claim 9 , wherein the computer is further configured to: obtain a plurality of training audio signals having one or more utterances of the one or more languages and a plurality of training labels corresponding to the plurality of training audio signals, each training label indicating the one or more languages in a corresponding audio training audio signal; extract a plurality of training embeddings by applying the embedding extractor extraction engine on the plurality of training audio signals; and train one or more layers of the embedding extractor extraction engine by applying loss function on the plurality of training embeddings and the plurality of training labels corresponding to the plurality of training audio signals.
13. The system according to claim 12 , wherein the computer is further configured to retrain the one or more layers of the embedding extractor extraction engine by applying the loss function on the cross-lingual quality measure, the enrolled voiceprint, and the speaker verification score.
14. The system according to claim 12 , wherein the computer is further configured to: select a subset of training audio signals having a plurality of languages of the plurality of training audio signals according to a subset of training labels of the plurality of training labels; and retrain the one or more layers of the embedding extractor extraction engine by applying the loss function on the subset of training embeddings and the subset of training labels.
15. The system according to claim 12 , wherein the computer is further configured to, for each training audio signal of the plurality of training audio signals: extract a plurality of features from the training audio signal; and generate an augmentation training signal corresponding to the training audio signal, the augmentation training signal comprising a plurality of reversed features corresponding to the plurality of features of the corresponding training audio signal.