简体中文 | English

ASRKit

One interface to run and compare any speech-to-text model — local & cloud.

一套接口，跑遍端云所有语音识别。

ASRKit is the Ollama + LiteLLM for speech recognition: pull and run open-source ASR models locally with one command, call cloud ASR APIs by changing one string, all through the same interface.

Install

pip install asrkit

One command sets up both local and cloud. Model weights are downloaded on demand via asrkit pull; cloud needs your API key (just like Ollama: install the engine, fetch models when you use them).

Quickstart

asrkit list                                              # list all models (✓ = installed)
asrkit run local/sensevoice a.wav                        # local: auto-download if missing, then transcribe
asrkit run siliconflow/sensevoice a.wav --api-key <KEY>  # cloud: just change the string

Python:

from asrkit import transcribe
r = transcribe("local/sensevoice", "a.wav")
print(r.text)

Highlights

• 47 on-device models (SenseVoice / Paraformer / Whisper / Zipformer / Moonshine / Parakeet / FireRed / Qwen3-ASR …), pull-and-go.
• OpenAI-compatible cloud endpoint, bring your own key (more providers landing).
• Transparent by design: never touches your audio or changes the model's native behavior by default; honest errors on format mismatch; conversion (--convert) and long-audio segmentation (--segment) are opt-in.
• Privacy: your audio and keys never pass through us — ASRKit runs on your machine.

Apache-2.0. Usage: docs/usage.md; write an adapter for a new engine/model: docs/adapter-spec.md.

Each model's license is defined by its own upstream (ASRKit is an interface and does not distribute weights; pull downloads from official releases); verify before commercial use. asrkit show <model> points you to the source.