🦜🔗 langchain-funasr

FunASR integration for LangChain — transcribe audio to LangChain Documents with self-hosted speech-to-text.

Powered by SenseVoice / Paraformer / Fun-ASR-Nano: runs locally, no cloud API, strong on Chinese and 50+ languages.

Install

pip install langchain-funasr

Usage

from langchain_funasr import FunASRLoader

loader = FunASRLoader("meeting.wav", model="iic/SenseVoiceSmall", device="cuda")
docs = loader.load()
print(docs[0].page_content)

Use the parser directly with blob pipelines:

from langchain_core.document_loaders import Blob
from langchain_funasr import FunASRParser

parser = FunASRParser(model="FunAudioLLM/SenseVoiceSmall", hub="hf", device="cuda")
docs = list(parser.lazy_parse(Blob.from_path("audio.wav")))

Options

Arg	Default	Notes
model	iic/SenseVoiceSmall	Any FunASR model (SenseVoice / Paraformer / Fun-ASR-Nano)
hub	ms	ms (ModelScope) or hf (HuggingFace)
device	cpu	e.g. cuda, cuda:0
language	auto	SenseVoice: auto/zh/en/yue/ja/ko
vad_model	fsmn-vad	Built-in VAD handles long audio of any length

Why FunASR

• Self-hosted — no API keys, no data leaving your machine.
• Fast — SenseVoice is non-autoregressive, far faster than Whisper.
• Strong on Chinese + 50+ languages.

⭐ If this helps, star FunASR and SenseVoice.

License

Apache-2.0