Data augmentation method based on stochastic feature mapping for automatic speech recognition
A method of augmenting training data includes converting a feature sequence of a source speaker determined from a plurality of utterances within a transcript to a feature sequence of a target speaker under the same transcript, training a speaker-dependent acoustic model for the target speaker for corresponding speaker-specific acoustic characteristics, estimating a mapping function between the feature sequence of the source speaker and the speaker-dependent acoustic model of the target speaker, and mapping each utterance from each speaker in a training set using the mapping function to multiple selected target speakers in the training set.
1. A method of augmenting training data comprising:
receiving, by a computer system, a training set of sampled audio data including speech of a plurality of source speakers;
for each of the plurality of source speakers;
converting, by the computer system, a feature sequence of a source speaker, of the plurality of source speakers, determined from a plurality of utterances of scripted speech within the training set, to a feature sequence of a respective target speaker under the same scripted speech, wherein the feature sequence of the respective target speaker is added to the training set;
training, by the computer system, a speaker-dependent acoustic model for the respective target speaker for corresponding speaker-specific acoustic characteristics; and
estimating, by the computer system, a mapping function between the feature sequence of the source speaker and the speaker-dependent acoustic model of the respective target speaker; and
for each of the plurality of source speakers:
mapping, by the computer system, each of the utterances from each of the plurality of source speakers in the training set using the mapping function to a plurality of other source speakers of the plurality of source speakers, wherein the mapping is added to the training set to generate augmented training data configured to train an automatic system recognition computer system.
2. The method of claim 1 , wherein the speaker-dependent acoustic model is estimated using a criteria including one or more of a maximum likelihood criterion, a maximum mutual information criterion and a minimum phone error criterion.
3. The method of claim 1 , wherein the mapping function is one of a linear function, an affine function and a nonlinear function.
4. The method of claim 1 , further comprising training the automatic system recognition computer system using the augmented training data.
5. The method of claim 1 , further comprising selecting the plurality of the source speakers randomly.
6. The method of claim 1 , further comprising selecting the plurality of the source speakers using a criterion including at least one of vocal tract length, dialect, and signal-to-noise ratio.
7. The method of claim 1 , wherein the respective target speaker is a new speaker created by perturbing a vocal tract length of the source speaker.