Self-Hosted & Alternative Transcription Providers

Voxis’s transcription client speaks the standard OpenAI-compatible /audio/transcriptions protocol: a multipart form (file, model, response_format=verbose_json, optional language, and optional task=translate) sent with an Authorization: Bearer <api_key> header, expecting back a JSON body with at least text. Any server implementing that same contract — another cloud provider, or a fully self-hosted Whisper-compatible server — works by pointing the app’s api_url_override at it. No code changes are required, and the app does not ship a provider-specific SDK per vendor or a platform-native transcription backend by design.

You can set this endpoint directly in the app: Settings → Provider → “Custom Endpoint URL (Advanced)” (stored as api_url_override), which validates the URL client-side and offers a one-click preset for the self-hosted speaches setup described below. The same field can also be set by tests or a custom build. See the full write-up linked below for exact steps, verified endpoints, and worked examples.

Quick start: self-hosted with Docker

This repository ships docker-compose.selfhost.yml at the project root — a thin wrapper around the upstream speaches image (OpenAI-compatible, backed by faster-whisper), with sane local defaults (binds to 127.0.0.1 only, persists downloaded models in a named volume).

docker compose -f docker-compose.selfhost.yml up -d
docker compose -f docker-compose.selfhost.yml exec speaches \
  curl -sX POST "http://localhost:8000/v1/models/Systran/faster-whisper-large-v3"

Then in Voxis Settings, set:

API URL override: http://localhost:8000/v1/audio/transcriptions
Model: Systran/faster-whisper-large-v3 (or whichever model you pulled)
API key: any non-empty placeholder — the server doesn’t check it, but the Voxis client always sends an Authorization: Bearer <key> header

A --profile gpu variant is included for NVIDIA GPU acceleration.

macOS-native option: Apple SpeechAnalyzer

On Apple Silicon Macs running macOS 26+, community projects (e.g. ohr) wrap Apple’s on-device SpeechAnalyzer/SpeechTranscriber API in an OpenAI-compatible HTTP server, so it plugs into api_url_override the same way as any other self-hosted option. It is fast and lightweight (verified ~10MB RSS growth in the calling process, out-of-process inference on the Neural Engine) and tested to be accurate for English, but it only supports 30 locales and does not support Russian (or most other non-English/es/fr/de/it/ja/ko/pt/ yue/zh languages) — for unsupported languages, use the Docker/speaches option above with a large-v3 model instead.

Full details

See docs/SELF_HOSTED_TRANSCRIPTION.md in the repository for:

The exact request/response protocol shape Voxis expects.
A table of known cloud endpoints (Groq default, OpenAI, Azure OpenAI).
Other protocol-compatible self-hosted options (whisper.cpp wrappers, LocalAI).
Verified memory, accuracy, and language-support results for both speaches and the macOS SpeechAnalyzer option, measured against this project’s real transcription client code.
Notes on servers with no authentication (the API key field still can’t be left empty).