talkis/docs/local-stt-runtime.md

Local STT Runtime

Talkis supports a managed local STT runtime named talkis-stt. The app uses it for local models without requiring Docker.

Runtime Lookup

When a local model is installed, Talkis:

1. Checks whether the local OpenAI-compatible endpoint is already available at http://127.0.0.1:8000.
2. Tries to start a bundled Tauri sidecar named talkis-stt.
3. If the sidecar is not bundled, downloads a platform-specific runtime from a manifest.
4. Starts the downloaded runtime and waits for /health.
5. Calls POST /v1/models/:model to install the selected model.

The runtime manifest URL defaults to:

https://talkis.ru/downloads/talkis-stt/manifest.json

For development it can be overridden with:

TALKIS_STT_RUNTIME_MANIFEST=https://example.test/manifest.json

Manifest Format

{
  "version": "0.1.0",
  "macos-aarch64": {
    "url": "https://talkis.ru/downloads/talkis-stt/0.1.0/talkis-stt-aarch64-apple-darwin",
    "sha256": "..."
  },
  "macos-x86_64": {
    "url": "https://talkis.ru/downloads/talkis-stt/0.1.0/talkis-stt-x86_64-apple-darwin",
    "sha256": "..."
  }
}

Runtime API

The runtime must bind to localhost only.

GET  /health
GET  /v1/models
POST /v1/models/:model
POST /v1/audio/transcriptions

GET /v1/models should return:

{
  "data": [
    { "id": "Systran/faster-whisper-large-v3-turbo" }
  ]
}

POST /v1/audio/transcriptions should accept OpenAI-compatible multipart form data and return:

{
  "text": "Recognized text"
}