简体中文 | English

ASRKit

One interface for speech recognition — cloud built in, engines on demand, models pull-and-go.

ASRKit is Ollama + LiteLLM for speech recognition: change one model string to swap between on-device models, cloud APIs, and any engine — your code doesn't change. The core is just a thin interface — cloud built in (tiny deps), local engines installed on demand, models pulled on demand, plus an optional OpenAI-compatible local server.

⚠️ Early beta, under active development. The core interface is usable, but we're still iterating — addressing/APIs may shift slightly between minor versions. Try it and tell us what breaks.

pip install asrkit                                        # seconds: interface + cloud (just requests)
asrkit transcribe audio.wav -m dashscope/qwen3-asr-flash --api-key <KEY>   # cloud, works instantly

pip install "asrkit[local]"                               # add on-device (sherpa, 47 models)
asrkit run local/sensevoice audio.wav                     # downloads on first run, then transcribes

Want a global command that doesn't touch your current env? pipx install asrkit or uv tool install asrkit work too.

One interface, swap anything

The model string is the only thing that changes — engine, model, on-device vs cloud, all the same call:

asrkit run local/sensevoice                      audio.wav              # sherpa-onnx (on-device)
asrkit run faster-whisper/large-v3               audio.wav              # faster-whisper engine
asrkit run whispercpp/base                       audio.wav              # whisper.cpp engine
asrkit run transformers/openai/whisper-large-v3  audio.wav              # any HuggingFace model
asrkit run dashscope/qwen3-asr-flash             audio.wav --api-key …  # cloud, bring your own key

from asrkit import transcribe
print(transcribe("dashscope/qwen3-asr-flash", "audio.wav", config={"api_key": "…"}).text)

Why ASRKit

• The interface is the core; everything is pluggable. pip install asrkit ships only the interface + cloud (tiny — just requests); engines, models, and the server are all add-ons. No torch for things you don't use.
• Same interface, local or cloud. On-device local/sensevoice and cloud siliconflow/sensevoice — swap the string, not your code.
• Curated and open. 47 on-device models + 5 cloud vendors out of the box; custom models are one JSON line, new engines are a plugin — no fork required.
• Transparent by design. It doesn't touch your audio or change a model's native behavior by default. Wrong format? An honest error — never silent garbage. Conversion and long-audio chunking are opt-in.
• Private. Your audio and API keys never leave your machine — it's a library/tool, not a hosted service.

Install: a tiny interface core, everything pluggable

The base install is just the interface + cloud (only requests — installs in seconds, runs anywhere). Add local engines on demand:

Use an engine you haven't installed → a friendly error with the install command, not an ImportError.

Ownership model: engines are shared pip packages — asrkit engine install <name> installs into the right environment for you; uninstall with your own pip uninstall (shared packages, your env, your call). Models are asrkit-owned — pull to download, rm to delete, clean and symmetric.

Commands

Engines & models

Cloud vendors (built in, just add a key)

Three ways to supply a key (highest priority first): --api-key > env var <VENDOR>_API_KEY (Volcengine's dual key uses DOUBAO_APP_KEY/DOUBAO_ACCESS_KEY) > asrkit config set-key <vendor> <KEY> once.

Run it as a server: OpenAI-compatible endpoint

asrkit serve starts a local server; any app on the OpenAI SDK (or an agent, or any language) reaches every model behind it by changing one base_url — this is the "LiteLLM proxy" half. The caller needs zero asrkit deps, just HTTP.

pip install "asrkit[serve]"
asrkit config set-key dashscope <KEY>     # store a key once (only for cloud)
asrkit serve                              # 127.0.0.1:11435 by default, local only

from openai import OpenAI
c = OpenAI(base_url="http://localhost:11435/v1", api_key="unused")
c.audio.transcriptions.create(model="local/sensevoice", file=open("a.wav", "rb"))

• Endpoints: POST /v1/audio/transcriptions (response_format: json/text/srt/vtt), GET /v1/models, GET /health.
• Cloud keys come from the asrkit config keystore — no per-request key needed. Transparent: raw bytes uploaded, no decoding.

Extend it

Add any sherpa model — one command, or an entry in ~/.asrkit/models.json:

asrkit add-model local/my-model --url https://…/model.tar.bz2 --arch senseVoice --langs zh,en

Add an engine — ship a small package that declares an asrkit.adapters entry point; pip install auto-registers it. No fork, no core changes. Recipe: docs/engines-and-addressing.md.

Docs

Usage · Adapter contract · Engines & addressing · Model management

Apache-2.0. Each model's license is its own — verify before commercial use (asrkit show <model> points to the source). Your audio and keys stay on your machine.