STT

STT - Local speech transcription sidecar

Run whisper.cpp locally on CPU or NVIDIA GPUs, with optional speaker labels from pyannote.

Given my privacy-consciousness, I wanted a speech-to-text service that could run on whatever machine I was using, from CPU-only Docker to an NVIDIA GPU. STT keeps whisper.cpp running behind a small HTTP API, so applications can send it audio without knowing how the model is deployed. I added separate Compose setups for regular Docker and rootless CDI, plus code that loads the selected model before startup. To annotate speaker labels (which are optional), a Python sidecar combines Whisper word timestamps with pyannote and groups the transcript into turns. I also measured the models on my own GPU so the repository says how much memory they need and how fast they run.

Open source project · 2023–now · Author