Verify in five minutes
Recommended path
Primary limitation
Open deployment guide
Open speech infrastructure
FunASR
Private-deployment speech infrastructure for high-throughput GPU inference, realtime streaming, OpenAI-compatible APIs, and standalone CPU or edge use.
- License
- Apache-2.0
- API
- OpenAI compatible
- Verified
- 2026-08-13
Deployment selector
Start with the workload, not a framework name
Choose the workload, hardware, and primary goal to get one recommendation with an explicit limitation.
Full matrix
The full decision path works without JavaScript
| Path | Workload | Hardware | Interface | Status |
|---|---|---|---|---|
| Native FunASR on vLLM | batch / private-api | nvidia-gpu / kubernetes | OpenAI-compatible HTTP | Community verified |
| SenseVoice TensorRT / Triton | batch / private-api | nvidia-gpu / kubernetes | Triton gRPC/HTTP / TensorRT plan | Production verified |
| llama.cpp / GGUF standalone | edge / batch | cpu / desktop-edge-gpu | CLI / local HTTP server | Production verified |
| SenseVoice native OpenAI realtime server | private-api / realtime / edge | cpu / nvidia-gpu | OpenAI-compatible HTTP / OpenAI realtime WebSocket | Production verified |
| audio.cpp native Fun-ASR-Nano and SenseVoice | edge / batch / private-api | cpu / nvidia-gpu / desktop-edge-gpu | CLI / OpenAI-compatible HTTP | Community verified |
| OpenAI-compatible private API | private-api / batch | cpu / nvidia-gpu / kubernetes | OpenAI-compatible HTTP / OpenAPI | Production verified |
| Realtime streaming and captions | realtime | cpu / nvidia-gpu / kubernetes | WebSocket / Python client | Community verified |
| Containers and Kubernetes | private-api / batch / realtime | kubernetes / cpu / nvidia-gpu | Docker Compose / Kubernetes ClusterIP | Community verified |
| ONNX / C++ CPU runtime | batch / realtime / edge | cpu | C++ / ONNX Runtime | Community verified |
| Production rollout checklist | private-api / batch / realtime / edge | cpu / nvidia-gpu / desktop-edge-gpu / kubernetes | HTTP / WebSocket | Production verified |
Inspectable API contract
Verify the service before integrating a client
Health, model identity, and real transcription are three separate checks.
$ curl -fsS http://localhost:8000/health
{"status":"ok"}
$ curl -fsS http://localhost:8000/v1/models
{"data":[{"id":"sensevoice"}]}
$ curl http://localhost:8000/v1/audio/transcriptions \
-F file=@meeting.wav -F model=sensevoiceProduction boundary
The runtime serves inference; the platform controls exposure
Native FunASR
Model loading, transcription, timestamps, VAD, streaming protocols, and runtime logs.
Gateway and platform
TLS, authentication, tenant quotas, rate limits, audit, upload policy, monitoring, and rollout.
Release evidence
Target-hardware capacity tests, known-audio smoke tests, version manifests, and executable rollback.
Ecosystem repositories
From models to video workflows
Open the focused repository for its models, deployment boundaries, and reproducible examples.
Production speech toolkit
FunASR
The open toolkit unifying ASR, VAD, punctuation, diarization, and production deployment.
Open FunASR on GitHub
High-accuracy ASR
Fun-ASR
A family of end-to-end speech recognition models spanning flagship and 31-language MLT checkpoints.
Open Fun-ASR on GitHub
Languages and emotion
SenseVoice
Multilingual understanding across speech recognition, language, emotion, and audio events.
Open SenseVoice on GitHub
AI video editing
FunClip
Video understanding and intelligent clipping workflows driven by speech and text localization.
Open FunClip on GitHub
Next step