Backends
Speaches
Speaches
is a lightweight transcription server that wraps Whisper-class models
behind an OpenAI-compatible /v1/audio/transcriptions
endpoint. Pendra uses it as the default audio backend and supports
one-click installs of Whisper-family catalogue models.
What's supported
| Capability | Status |
|---|---|
| Chat completions | — |
| Embeddings | — |
| Image generation | — |
| Audio transcription | ✓ |
| Model install | ✓ |
| Model uninstall | ✓ |
Connection
Speaches is discovered via the same probe-then-verify flow as the other backends. The verification step confirms the server speaks the Speaches API. Override is the same as any other backend — set its endpoint env var in worker config.
Model install & uninstall
Whisper variants in the Pendra catalogue (e.g.
whisper-large-v3-turbo, whisper-tiny.en) can be
installed via the console. Behind the scenes Pendra issues a
POST /v1/models/<id> call against Speaches, which
downloads and registers the model. Progress streams live the same way as
Ollama installs.
Uninstall works the same way — click Remove on an
installed model and Pendra calls
DELETE /v1/models/<id> on the worker's Speaches
instance.