Noise adaptive training for speech recognition
Technologies are described herein for noise adaptive training to achieve robust automatic speech recognition. Through the use of these technologies, a noise adaptive training (NAT) approach may use both clean and corrupted speech for training. The NAT approach may normalize the environmental distortion as part of the model training. A set of underlying “pseudo-clean” model parameters may be estimated directly. This may be done without point estimation of clean speech features as an intermediate step. The pseudo-clean model parameters learned from the NAT technique may be used with a Vector Taylor Series (VTS) adaptation. Such adaptation may support decoding noisy utterances during the operating phase of a automatic voice recognition system.
1. A computer-implemented method for speech recognition, the method comprising computer-implemented operations for:
receiving a set of training speech utterances, wherein one or more of the training speech utterances comprises speech corrupted by noise or distortion;
initializing parameters of a pseudo-clean speech model;
initializing parameters of a distortion model by setting a channel mean to zero and setting a noise mean and a covariance according to one of a first or a last sample of an utterance where no speech is present;
iteratively applying noise adaptive training using the set of training speech utterances to generate a speech recognition model based upon identifying a set of distortion model parameters from the distortion model and identifying a set of pseudo-clean model parameters from the pseudo-clean speech model;
receiving a speech signal to be recognized;
adapting the speech recognition model to a current environment associated with the received speech signal; and
performing speech recognition on the received speech signal using the adapted speech recognition model.
2. The computer-implemented method of claim 1 , wherein a first one of the training speech utterances comprises a different noise or distortion than a second one of the training speech utterances.
3. The computer-implemented method of claim 1 , wherein the speech recognition model comprises a Hidden Markov Model.
4. The computer-implemented method of claim 1 , wherein generating the speech recognition model further comprises incorporating a Vector Taylor Series adaptation into training of the speech recognition model.
5. The computer-implemented method of claim 1 , wherein generating the speech recognition model further comprises iterating over the training speech utterances while accumulating statistics to update the speech recognition model.
6. The computer-implemented method of claim 1 , wherein generating the pseudo-clean model comprises approximating a model trained with clean speech.
7. A computer system comprising:
a processing unit;
a memory operatively coupled to the processing unit; and
a program module which executes in the processing unit from the memory and which, when executed by the processing unit, causes the computer system to recognize speech by
receiving a set of training speech utterances, wherein one or more of the training speech utterances comprises speech corrupted by noise or distortion;
initializing parameters of a pseudo-clean speech model;
initializing parameters of a distortion model by setting a channel mean to zero and setting a noise mean and a covariance according to one of a first or a last sample of an utterance where no speech is present;
iteratively applying noise adaptive training using the set of training speech utterances to generate a speech recognition model based upon identifying a set of distortion model parameters from the distortion model and identifying a set of pseudo-clean model parameters from the pseudo-clean model;
receiving a speech signal to be recognized;
adapting the speech recognition model to a current environment associated with the received speech signal; and
recognizing speech elements from the received speech signal using the adapted speech recognition model.
8. The computer system of claim 7 , wherein a first one of the training speech utterances comprises a different noise or distortion than a second one of the training speech utterances.
9. The computer system of claim 7 , wherein the speech recognition model comprises a Hidden Markov Model.
10. The computer system of claim 7 , wherein generating the speech recognition model comprises incorporating a Vector Taylor Series adaptation into training of the speech recognition model.
11. The computer system of claim 7 , wherein generating the speech recognition model comprises iterating over the training speech utterances while accumulating statistics to update the speech recognition model.
12. A computer-readable medium having computer-executable instructions stored thereon which, when executed by a computer, cause the computer to:
receive a set of training speech utterances, wherein at least one of the training speech utterances comprises speech corrupted by noise or distortion;
initializing parameters of a hidden Markov speech recognition model;
initializing parameters of a distortion model by setting a channel mean to zero and setting a noise mean and a covariance according to one of a first or a last sample of an utterance where no speech is present;
generate an improved hidden Markov speech recognition model using the set of training speech utterances by applying a Vector Taylor Series adaptation to training of the hidden Markov speech recognition model, and generating a set of pseudo-clean model parameters from the hidden Markov speech recognition model and a set of distortion model parameters from the distortion model; and
recognize speech elements using the improved hidden Markov speech recognition model.