ScribeCrate

ScribeCrate is an open-source, MIT-licensed transcription API powered by faster-whisper. It exposes OpenAI Whisper-compatible endpoints, runs in Docker on CPU or NVIDIA GPU, and keeps audio on the user's own server.

ScribeCrate is an open-source transcription API that runs on hardware the user controls. It wraps faster-whisper and Whisper models behind endpoints modeled on the OpenAI Whisper API, so recordings are converted to text, subtitles or English translations without leaving the server. The project is MIT licensed, and the models and bundled dependencies keep their own licenses. The tool is aimed at developers, self-hosting enthusiasts and teams that handle sensitive audio such as meeting recordings, interviews and podcasts. Any client that already speaks the Whisper API can point at the /v1/audio/transcriptions or /v1/audio/translations endpoint instead of a hosted service, and no OpenAI account or key is needed. Output can be JSON, plain text, verbose JSON, SRT or WebVTT. Uploaded files can also be streamed back segment by segment over SSE. An optional local sherpa-onnx extension adds speaker diarization, which labels who is talking in each segment, though it does not apply in streaming mode. In practice, deployment is a single Docker container. The base model downloads on first start, an API key is generated automatically on a fresh install, and a curl call with an audio file returns the transcript. The default setup needs roughly 700 MB of free RAM. Images exist for amd64 and arm64 CPUs, and a :cuda image gives NVIDIA acceleration on amd64. Model size ranges from tiny to large-v3 and large-v3-turbo, and anything FFmpeg can decode is accepted as input. With models cached in advance and a local-only setting enabled, the server runs fully offline. Among alternatives, ScribeCrate sits between paid cloud transcription services and hand-assembled faster-whisper scripts. It trades convenience for control: the user supplies the hardware, handles HTTPS and authentication when exposing the API, and does the upkeep. It handles uploaded files only; live microphone work belongs to the separate ScribeCrate Live project. A Colab notebook offers a no-install trial.