Re:Firefox Voice
by nshm on Monday August 10, 2020 @03:56PM
(#24113350)
You have wrong premise here. First of all, current Mozilla data is almost useless for training because it is carefully read speech. You do not need much of it, even accents. If you add 1000 hours of CV data to 1000 hours of random data the improvement in accuracy will be minimal. Same for the speech collected with Firefox Voice, it will be mostly a set of short commands, most likely it will not be very useful for generic transcription of random people speech.
Second, you can build models much better than Mozilla ones simply from public data, there is no need to collect user voices. We at Vosk https://alphacephei.com/vosk/ support 10+ languages for example without any user data. Everyone creates very good models from augmented text-to-speech data these days (Microsoft demonstrated in the last paper you can get almost as good as domain-specific data https://arxiv.org/abs/2007.15188).
Given that it surprises me that Mozilla continues to insist they need the voices of their userbase.
Second, you can build models much better than Mozilla ones simply from public data, there is no need to collect user voices. We at Vosk https://alphacephei.com/vosk/ support 10+ languages for example without any user data. Everyone creates very good models from augmented text-to-speech data these days (Microsoft demonstrated in the last paper you can get almost as good as domain-specific data https://arxiv.org/abs/2007.15188).
Given that it surprises me that Mozilla continues to insist they need the voices of their userbase.