5.2.48. Speaches (unified TTS + STT engine)

Speaches is an STT and TTS engine. One container serves /v1/audio/transcriptions (Faster-Whisper) and /v1/audio/speech (Kokoro and Piper voices). Select it with STT_PROVIDER_SOURCE=speaches-* or TTS_PROVIDER_SOURCE=speaches-*. When both select it, one container serves both.

Atlas preloads no model, and Speaches does not download a model on request. Requests return 404 until you download the model (issue #799). See the STT Provider README for STT and the TTS Provider README for TTS.

1. Engine quick reference

• Images:
  • CPU: ghcr.io/speaches-ai/speaches:0.9.0-rc.3-cpu
  • GPU: ghcr.io/speaches-ai/speaches:0.9.0-rc.3-cuda
• License: MIT
• Activation: any of
  • STT_PROVIDER_SOURCE=speaches-container-cpu
  • STT_PROVIDER_SOURCE=speaches-container-gpu
  • TTS_PROVIDER_SOURCE=speaches-container-cpu
  • TTS_PROVIDER_SOURCE=speaches-container-gpu
• GPU caveat: speaches-container-gpu selects the CUDA image, but the service gets no NVIDIA device (issue #1373, open). It runs on CPU or fails CUDA initialisation. Use speaches-container-cpu.
• In-container port: 8000
• Host port: ${SPEACHES_PORT} (computed from BASE_PORT)

2. Capabilities & limitations

Support tier: experimental — Capability contract declared; no cited cold-start, workflow, or upgrade qualification run yet (evidence at v0.1.0).

Capability	Status	Verification	Notes
OpenAI-compatible text-to-speech	partial	tested	Speaches serves /v1/audio/speech, but Atlas does not preload Kokoro; the model must be downloaded before requests succeed.
OpenAI-compatible speech-to-text	partial	untested	Speaches exposes /v1/audio/transcriptions, but Atlas has not validated the current preload and Open WebUI model path against a live container.
Configurable STT model selection	stubbed	documented	SPEACHES_STT_MODEL is declared but does not alter the hard-coded PRELOAD_MODELS value or Open WebUI's STT model.

3. Dependencies & Integrations

3.1. Current — Upstream (this service calls)

No upstream calls.

3.2. Current — Downstream (services that call this)

Service	Category
kong	infra
hermes	agents
n8n	agents
jupyterhub	apps
open-webui	apps

3.3. Architecture diagram

Open the full-size diagram for a full-screen view.

3.4. Future — Missing pair integrations

No high-confidence opportunities identified.

3.5. Future — Candidate new services

No high-confidence opportunities identified.

3.6. Future — Unused features in this service

No high-confidence opportunities identified.