Re: legal questions regarding machine learning models

[Mathieu Blondel <mathieu@mblondel.org>]

Replying to Paul Wise (sorry I'm not subscribed to the mailing-list, I
saw your message through the archive)

Modern speech recognition engines are usually speaker independent. In
order to support speaker dependent models, users would have to record
their voice in order to train the models. This may be fine for small
vocabularies but this is not feasible for large vocabularies.
Moreover, there exists techniques to adapt the speaker-independent
models to one speaker. These techniques require much less training
data than creating a speaker dependent model from scratch would
require. And since humans are able to recognize many different voices,
shouldn't we have the same goal for speech recognition?

Thank you,
Mathieu Blondel