Method and system for acoustic model conditioning on non-phoneme information features
A method and system for acoustic model conditioning on non-phoneme information features for optimized automatic speech recognition is provided. The method includes using an encoder model to encode sound embedding from a known key phrase of speech and conditioning an acoustic model with the sound embedding to optimize its performance in inferring the probabilities of phonemes in the speech. The sound embedding can comprise non-phoneme information related to the key phrase and the following utterance. Further, the encoder model and the acoustic model can be neural networks that are jointly trained with audio data.
1. A computer-implemented method of inferring phoneme probabilities in speech audio, the method comprising:
receiving, at a computing device, a first segment of speech from a user, wherein the first segment of speech corresponds to a key phrase with known phonemes;
generating, using an encoder model, a sound embedding from the first segment of speech, wherein the sound embedding comprises information of non-phoneme features of the first segment of speech, wherein the information of non-phoneme features comprises at least one of voice, noise, accent, and environmental attributes associated with the first segment of speech;
receiving a second segment of speech following the first segment of speech from the same user; and
inferring, using an acoustic model conditioned on the sound embedding, the phoneme probabilities related to the second segment of speech.
2. The computer-implemented method of claim 1 , further comprising:
storing the sound embedding in a memory device associated with the computing device.
3. The computer-implemented method of claim 1 , wherein the acoustic model is trained on labeled samples of speech audio, each of the labeled samples having a corresponding sound embedding.
4. The computer-implemented method of claim 3 , wherein the labeled samples include a multiplicity of voices mixed with a multiplicity of noise profiles, wherein the first segment and the second segment are mixed with the same noise profile for each sample.
5. The computer-implemented method of claim 1 , wherein the encoder model is jointly trained with the acoustic model.
6. The computer-implemented method of claim 1 , further comprising:
generating, using a second encoder model, a second sound embedding from the first segment of the speech, wherein the acoustic model is further conditioned on the second sound embedding to infer the phoneme probabilities related to the second segment of speech.
7. A computer-implemented method of conditioning an acoustic model for speech recognition, the method comprising:
receiving, at an acoustic model, a sound embedding comprising information of non-phoneme features of a first segment of speech from a user, wherein the first segment of speech corresponds to a key phrase with known phonemes, wherein the acoustic model is trained to condition its output based on the sound embedding, wherein the information of non-phoneme features comprises at least one of voice, noise, accent, and environmental attributes associated with the first segment of speech;
receiving a second segment of speech following the first segment of speech from the same user; and
inferring, using the acoustic model, phoneme probabilities of the second segment of speech.
8. The computer-implemented method of claim 7 , further comprising:
generating, using an encoder model, the sound embedding based on an audio signal of the first segment of speech, wherein the first segment of speech corresponds to a key phrase with known phonemes.
9. The computer-implemented method of claim 8 , further comprising:
generating, using a second encoder model, a second sound embedding from the first segment of speech, wherein the acoustic model is further trained on the second sound embedding to infer the phoneme probabilities related to the second segment of speech.
10. The computer-implemented method of claim 8 , further comprising:
training the encoder model jointly with the acoustic model based on the sound embedding.
11. The computer-implemented method of claim 7 , wherein the acoustic model is trained on labeled samples of speech audio, each of the labeled samples having a corresponding sound embedding.
12. The computer-implemented method of claim 11 , wherein the labeled samples include a multiplicity of voices mixed with a multiplicity of noise profiles, wherein the first segment and the second segment are mixed with the same noise profile for each sample.
13. A computerized speech processing system, the system being configured to:
receive an audio signal of a key phrase from a user, wherein the key phrase has known phonemes;
generate, using an encoder model, a sound embedding based on the audio signal, wherein the sound embedding comprises information of non-phoneme features of the key phrase, wherein the information of non-phoneme features comprises at least one of voice, noise, accent, and environmental attributes associated with the first segment of speech;
receiving speech following the key phrase from the same user; and
recognize the speech using an acoustic model, wherein the acoustic model is trained to condition its output based on the sound embedding.
14. The computerized speech processing system of claim 13 , wherein the outputs of the acoustic model comprise phoneme probabilities of an utterance following the key phrase.
15. The computerized speech processing system of claim 13 , wherein the information of non-phoneme features comprise at least one of voice, noise, accent, and environmental attributes associated with the key phrase.
16. The computerized speech processing system of claim 13 , wherein the encoder model is jointly trained with the acoustic model based on the sound embedding.
17. The computerized speech processing system of claim 13 , further configured to:
store the sound embedding in a memory device; and
transmit the stored sound embedding and an audio signal of an utterance following the key phrase over a network to a speech recognition server.
18. The computerized speech processing system of claim 13 , further configured to:
receive an audio signal of an utterance following the key phrase; and
infer, using the trained acoustic model, phoneme probabilities of the utterance.
19. The computerized speech processing system of claim 13 , further configured to:
generate, using a second encoder model, a second sound embedding based on the audio signal, wherein the acoustic model is trained to condition its output based on the second sound embedding.
20. The computerized speech processing system of claim 13 , wherein the acoustic model is trained on labeled samples of speech audio, each of the labeled samples having a corresponding sound embedding, and
wherein the labeled samples include a multiplicity of voices mixed with a multiplicity of noise profiles, wherein the first segment and the second segment are mixed with the same noise profile for each sample.