5.2.41. Parakeet (STT engine)

Parakeet is an STT engine selected with STT_PROVIDER_SOURCE. See the STT Provider README for setup and configuration.

1. Engine quick reference

• Image: built locally from provider/gpu/Dockerfile on nvcr.io/nvidia/pytorch:26.06-py3 (GPU only; Parakeet-TDT model)
• License: model CC-BY-4.0; container base NVIDIA-DLC
• Activation: STT_PROVIDER_SOURCE=parakeet-container-gpu (or parakeet-localhost — Parakeet-MLX on macOS / native Linux)
• In-container port: 8000
• Host port: ${STT_PROVIDER_PORT} (computed from BASE_PORT by the bootstrapper)
• Readiness: Uvicorn starts while a deadline-bounded background task loads the model. GET /health returns 503 until loading completes. A cold first boot downloads the ~2.5 GB checkpoint. The healthcheck start period (810 s plus retries) covers the 900 s load deadline.

This manifest also owns the STT_PROVIDER_SOURCE list for all STT engines.

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
NVIDIA Parakeet transcription	supported	tested	The GPU source loads the configured Parakeet-TDT checkpoint and serves standard and advanced OpenAI-shaped transcription routes with truthful readiness.
Host speech-to-text variants	partial	tested	Atlas can route to operator-run Parakeet MLX or whisper.cpp endpoints, but installation, model provisioning, process supervision, and host acceleration remain operator-owned.
Timestamp-rich transcription	partial	tested	Atlas-managed Parakeet providers expose segment and word timing when the selected implementation returns alignment data; plain results honestly report timestamps unavailable.
Provider authentication and workload bounds	partial	tested	Atlas Parakeet requires a bearer token and bounds uploads, admission, and inference by default; those guarantees do not extend to selected Speaches or whisper.cpp upstreams.
Cross-architecture Parakeet container	not-supported	tested	The only Parakeet container is NVIDIA GPU. Apple Silicon uses the separately installed MLX localhost provider; no CPU container or GPU quantization knob is advertised.

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.