Language
Yoruba Live Voice
An always-on Yoruba speech pipeline that never stops listening.
Speak Yoruba into a microphone and the system either captions what you said with correct accents, or answers you as a voice assistant and speaks the reply aloud.
The engineering problem is that cloud calls are slow and microphones are not patient. Five stages run concurrently so a slow network round trip never blocks capture or live decoding — voice activity detection and speech-to-text stay local, and only correction and speech synthesis leave the machine.
How it is built
- Five concurrent stages, so the microphone is never deaf while a slow stage is working.
- Streaming voice-activity detection emits live partial snapshots and finalises on a pause.
- Speech-to-text runs locally; only the correction and synthesis steps are remote.
- Two modes from the same pipeline: accented captioning, or spoken conversation.
- Now drives hardware: wired to an ESP32-S3 with motor drivers, spoken Yoruba becomes movement with no English step in between.