Speech & communicationConstructionism & lived experienceSpeech recognition & AI
AVoz: DIY speech recognition
2026
Speech recognition trained to understand my own voice, and AVoz, a mobile app that turns what I say into a clear synthesised voice.
- The challenge
- Dictation, automatic captions and voice assistants are trained on typical speech. For people with dysarthria, like me, they make too many mistakes to be useful.
- The approach
- Fernando Lobo and I fine-tuned OpenAI’s Whisper models (medium and large-v3) with LoRA on recordings of my own speech, using consumer-grade hardware. AVoz then speaks the recognised text with a synthesised voice, with all processing done locally.
- Outcomes
- On 100 held-out utterances, the word error rate fell from 49% to 17% with the medium model and from 30% to 12% with large-v3. The method is replicable: the code, the LoRA configuration and a step-by-step guide are open source (MIT licence), so other people can train a model for their own voice, in their own language.
Preprint, submitted to a peer-reviewed journal. It has not been peer-reviewed yet.
Related publications
- Condado, P. A., & Lobo, F. G. (2026). DIY customized speech recognition for a person with speech disabilities (arXiv:2611.11327). arXiv. https://arxiv.org/abs/2611.11327