Common Voice

And so too with openness.

And so you have to look at, like, is openness actually serving the terminal value or not, right? And so what that means for Mozilla is just earlier this year, recognizing this concern around data, we were like, you know, we've been running Common Voice, which is like the world's largest crowdsourced voice data set.

We have a number of indigenous communities and other people who are stewards of different languages saying, hey, I don't love How a open data set of my language is being used to create deepfake scam calls for my elders.

So I want to change the terms of its openness.

The audio is fussy, too, in a more mechanical way.

Take the standard Hugging Face walkthrough for fine-tuning Whisper on Common Voice.

The source clips are 48 kHz.

Whisper's feature extractor wants 16.

Nothing in the pipeline complains.

I'm fairly sure of one real case, a Thai speech model.

A wave to VEC2, trained on common voice, where the default split shared speakers between train and test, and the authors re-split the corpus before training.

The default was leaking, and a lot of people had presumably been reporting scores off it.

If your smallest source exhausts early and you stop, you never see the big one again.

There's a work notebook in Hugging Face's community events repo that builds a Spanish blend out of Common Voice, Vox Populi, multilingual LibreSpeech, and FLIRS.

The stated motivation is exactly the low resource case.

Any one of those alone might not have enough data, so you interleave all four.

They fine-tuned Whisper on a mixed Cantonese-English dataset they generated themselves using a multi-agent framework, and the media and mini datasets came in from the HKUST collection we talked about.

The headline is fourteen point two eight percent mix error rate.

That's thirty-five percent lower than the original Whisper on the same task, and twelve point six one percent character error rate on Common Voice Cantonese.

So a model that was never trained for the mix, plus a purpose-built blend, plus one training run.

Drastic.

Take Whisper Small on Divehi from Common Voice.

Orthographic WEA SA hundred and sixty-seven percent.

Run it through the normalizer and it drops to a hundred and twenty-six.

LibriSpeech is audiobook-read speech, and it's everybody's favorite because the audio is clean and the speakers are cooperative.

FLRS is Wikipedia sentences read by native speakers, a hundred and two languages.

Common Voice is crowdsourced volunteer recordings, so quality varies by contributor.

Multilingual LibriSpeech.