Age estimation from speech
Disclosed are systems and methods including computing-processes executing machine-learning architectures implementing label distribution loss functions to improve age estimation performance and generalization. The machine-learning architecture includes a front-end neural network architecture defining a speaker embedding extraction engine of the machine-learning architecture, and a backend neural network architecture defining an age estimation engine of the machine-learning architecture. The embedding extractor is trained to extract low-level acoustic features of a speaker's speech, such as mel-frequency cepstral coefficients (MFCCs), from audio signals, and then extract a feature vector or speaker embedding vector that mathematically represents the low-level features of the speaker. The age estimator is trained to generate an estimated age for the speaker and a Gaussian probability distribution around the estimated age, by applying the various types of layers of the age estimator on the speaker embedding.
1. A computer-implemented method comprising:
obtaining, by a computer, a plurality of training speaker embeddings extracted for a plurality of training audio signals of a plurality of training speakers;
training, by the computer, an age estimator of a machine-learning architecture to generate an estimated age and an estimated probability distribution by applying the age estimator on the plurality of training speaker embeddings of the plurality of training speakers using a plurality of speaker labels, each speaker label indicating an actual age of a corresponding training speaker, wherein the computer trains the age estimator based upon comparing a predicted probability distribution against an expected probability distribution;
obtaining, by the computer, an inbound speaker embedding for an inbound audio signal of an inbound speaker;
generating, by the computer applying the age estimator on the inbound speaker embedding, an inbound probability distribution for the inbound speaker; and
generating, by the computer applying the age estimator on the inbound speaker embedding, the estimated age of the inbound speaker and a confidence score based upon the estimated inbound probability distribution.
2. The method according to claim 1 , wherein training the age estimator includes:
for each training signal, generating, by the computer, a predicted estimated age based upon applying the age estimator on a training embedding of the training signal of a training speaker; and
determining, by the computer, a loss based on a difference between the predicted estimated age and the actual age of the training speaker of the training signal according to the speaker label.
3. The method according to claim 2 , further comprising updating, by the computer, one or more weights of the age estimator based upon the loss.
4. The method according to claim 2 , wherein generating the predicted estimated age includes generating, by the computer, the predicted probability distribution for the predicted estimated age of the training signal, wherein the predicted estimated age is a mean of the predicted probability distribution; and
wherein determining the difference includes determining, by the computer, a distance between the predicted probability distribution and the expected probability distribution, wherein the computer updates the expected probability distribution based upon the distance.
5. The method according to claim 1 , further comprising determining, by the computer, that the age estimator is trained in response to determining that a difference between the predicted probability distribution and the expected probability distribution satisfies a training threshold.
6. The method according to claim 1 , further comprising updating, by the computer, one or more weights of the age estimator based upon a difference between the inbound probability distribution and the expected probability distribution to update training of the age estimator.
7. The method according to claim 1 , further comprising receiving, by the computer, from a database the actual age of a speaker label of a training speaker.
8. The method according to claim 1 , wherein obtaining a training speaker embedding of a training audio signal includes:
applying, by the computer, an embedding extractor of the machine-learning architecture on the training audio signal to extract a plurality of acoustic features for the training audio signal; and
applying, by the computer, the embedding extractor on the plurality of acoustic features to extract the training speaker embedding for the training audio signal.
9. The method according to claim 1 , further comprising:
responsive to the computer determining that the inbound speaker embedding satisfies a speaker-match threshold distance from an enrolled speaker embedding, and that the estimated age is in an age-match threshold distance from an enrolled age of the enrolled speaker:
identifying, by the computer, the inbound speaker as the enrolled speaker, in response to the computer.
10. The method according to claim 1 , further comprising receiving, by the computer, one or more parameters of the expected probability distribution from a graphical user interface of an administrator device, wherein the computer determines the expected probability distribution based upon the one or more parameters.
11. A system comprising:
a computer comprising a processor configured to:
obtain a plurality of training speaker embeddings extracted for a plurality of training audio signals of a plurality of training speakers;
train an age estimator of a machine-learning architecture to generate an estimated age and an estimated probability distribution by applying the age estimator on the plurality of training speaker embeddings of the plurality of training speakers using a plurality of speaker labels, each speaker label indicating an actual age of a corresponding training speaker, wherein the computer trains the age estimator based upon comparing a predicted probability distribution against an expected probability distribution;
obtain an inbound speaker embedding for an inbound audio signal of an inbound speaker;
generate, by applying the age estimator on the inbound speaker embedding, an inbound probability distribution for the inbound speaker; and
generate, by applying the age estimator on the inbound speaker embedding, the estimated age of the inbound speaker and a confidence score based upon the estimated inbound probability distribution.
12. The system according to claim 11 , wherein, when training the age estimator, the computer is further configured to:
for each training signal, generate a predicted age based upon applying the age estimator on a training embedding of the training signal of a training speaker; and
determine a loss based on a difference between the predicted estimated age and the actual age of the training speaker of the training signal according to the speaker label.
13. The system according to claim 12 , wherein the computer is further configured to update one or more weights of the age estimator based upon the loss.
14. The system according to claim 12 , wherein, when generating the predicted estimated age, the computer is further configured to generate the predicted probability distribution for the predicted age of the training signal, wherein the predicted age is a mean of the predicted probability distribution; and
wherein, when determining the difference, the computer is further configured to determine a distance between the predicted probability distribution and the expected probability distribution, wherein the computer updates the expected probability distribution based upon the distance.
15. The system according to claim 11 , wherein the computer is further configured to determine that the age estimator is trained, in response to the computer determining that a difference between the predicted probability distribution and the expected probability distribution satisfies a training threshold.
16. The system according to claim 11 , wherein the computer is further configured to update one or more weights of the age estimator based upon a difference between the inbound probability distribution and the expected probability distribution to update training of the age estimator.
17. The system according to claim 11 , the computer is further configured to receive from a database the actual age of a speaker label of a training speaker.
18. The system according to claim 11 , wherein, when obtaining a training speaker embedding of a training audio signal, the computer is further configured to:
apply an embedding extractor of the machine-learning architecture on the training audio signal to extract a plurality of acoustic features for the training audio signal; and
apply the embedding extractor on the plurality of acoustic features to extract the training speaker embedding for the training audio signal.
19. The system according to claim 11 , wherein the computer is further configured to:
responsive to the computer determining that the inbound speaker embedding satisfies a speaker-match threshold distance from an enrolled speaker embedding, and that the estimated age is in an age-match threshold distance from an enrolled age of the enrolled speaker:
identify the inbound speaker as the enrolled speaker.
20. The system according to claim 11 , wherein the computer is further configured to receive one or more parameters of the expected probability distribution from a graphical user interface of an administrator device, wherein the computer determines the expected probability distribution based upon the one or more parameters.