MIYUPPIE/STT

Always-on Yoruba speech pipeline. Speak Yoruba into your mic and the system either: captions - what you said and prints a corrected, properly accented version (default), or answers - you as a conversational voice assistant and speaks the reply aloud. # Yoruba Live Voice Always-on **Yoruba** speech pipeline. Speak Yoruba into your mic and the system either: - **captions** what you said and prints a corrected, properly accented version (default), or - **answers** you as a conversational voice assistant and speaks the reply aloud (`--chat`). Whisper (STT) and Silero (VAD) run **locally**. Grok (xAI) does the correction/conversation and YarnGPT does the Yoruba speech — both are cloud APIs keyed from `.env`. ``` mic ─► Silero VAD ─► faster-whisper (STT) ─► Grok (correct | respond) ─► YarnGPT (TTS) ─► speaker local local cloud cloud ``` ## What it does `live_caption.py` runs five concurrent stages so the mic is never deaf — slow cloud calls never block capture or live decoding: 1. **PortAudio callback** pushes 512-sample frames into a queue. 2. **Streaming Silero VAD** finds speech start/end, emitting live partial snapshots and finalized segments. 3. **faster-whisper** decodes each segment, printing partials live (`\r`) and finalizing on a pause. 4. The **Grok** stage either **corrects** the caption (default refiner) or **responds** to it as a conversational brain (`--chat`). 5. **YarnGPT** reads the correction/reply aloud (`--speak` / `--chat`). Speaking is **half-duplex**: the mic mutes during playback so the spoken audio isn't transcribed back into a feedback loop. ## Repo layout ``` live_caption.py # the always-on app (all five stages) benchmark.py # transformers fp32 vs faster-whisper fp16 on one utterance download.py # pull the base Whisper Yoruba model from Hugging Face test.py # standalone record-until-silence STT loop (faster-whisper) services/ brain/ # Grok conversational brain (--chat) — responds refine/ # Grok caption corrector (default) — corrects tts/ # YarnGPT Yoruba text-to-speech env_loader.py # loads repo-root .env into …