1
$\begingroup$

My team is building a speech recognition model. As far as I know, Vocal Tract Length Normalization (VTLN) is an effective warping method that may help improve the model performance by diminishing spectral variance.

However, most implementations I've seen so far seem to use maximum-likelihood to estimate the warping factor, which could be slow when the model is used to recognize real-time audio streams (we plan to deploy the model on hardware so some computations in calculating probability such as division may be expensive).

Is there a fast implementation of VTLN? It's okay if the performance is slightly worse than the ML implementation.

Improve this question
$\endgroup$
1
  • $\begingroup$ Why are you worried about the complexity of computation already? Do you have a thing that works offline but wouldn't work in real-time? $\endgroup$
    – Marcus Müller
    Commented Jun 11, 2019 at 11:49

1 Answer 1

Reset to default
1
$\begingroup$

VTLN is pretty old technology. There is a fast linear VTLN or basis fMLLR which is about the same thing. These days everyone is doing neural networks though, they provide more advanced speaker adaptation, at the same time they could be pretty efficient with quantization and proper weight pruning.

For example you can check DOMAIN AND SPEAKER ADAPTATION FOR CORTANA SPEECH RECOGNITION by Microsoft.

Best adaptation method depend more on the task you are trying to solve, on the resource constrains you have, on the model you are using.

Improve this answer
$\endgroup$
2
  • $\begingroup$ Thanks for your answer. It seems the paper illustrates some speaker adaption methods to DNN. Do these methods also apply to CNN as well? $\endgroup$
    – Steven Chan
    Commented Jun 12, 2019 at 4:51
  • $\begingroup$ CNN is a kind of DNN architecture, so yes, you can use similar things. $\endgroup$
    – Nikolay Shmyrev
    Commented Jun 12, 2019 at 8:18

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.