New project · Preprint
AVoz: speech recognition that understands my voice
Most speech recognition is trained on typical voices and fails on mine. With Fernando Lobo, I fine-tuned an open speech model on recordings of my own speech, using consumer-grade hardware, and built AVoz, an app that turns what I say into a clear synthesised voice. The code is open source, so other people can do the same for their own voice.
- Word error rate
- 30% down to 12%
- Character error rate
- 16% down to 5%
Whisper large-v3 before and after fine-tuning, on 100 held-out utterances. Preprint submitted to a peer-reviewed journal; not yet peer-reviewed.
