FunASR Speech Recognition Plugin

Self-hosted speech recognition through FunASR's OpenAI-compatible API. The
plugin supports multilingual ASR models for private Dify deployments without
sending audio to a hosted transcription service.

In FunASR's
192-minute benchmark,
SenseVoiceSmall reached 170x real-time on GPU and 17x real-time on CPU.
Whisper-large-v3 reached 13x real-time on GPU in the same benchmark. Results
depend on the hardware, audio, model, and concurrency settings.

Setup

Deploy a SenseVoice server:

Use when CUDA is unavailable. Fun-ASR-Nano deployments use
vLLM and can be installed separately:

In Dify, configure this plugin with the server URL, for example
or . The plugin
normalizes both forms to a base URL ending in , such as
. Leave the API key empty when the server does
not require one.

Predefined models accept audio files up to 25 MB.

Supported Models

sensevoice - Mandarin, Cantonese, English, Japanese, and Korean ASR with
emotion and audio-event tags (default)
paraformer - Chinese ASR with punctuation
paraformer-en - English ASR
fun-asr-nano - Chinese, English, Japanese, and Chinese dialect/accent ASR
with an encoder + LLM architecture

FunASR Speech Recognition Plugin

Setup

Deploy a SenseVoice server:

Use when CUDA is unavailable. Fun-ASR-Nano deployments use
vLLM and can be installed separately:

In Dify, configure this plugin with the server URL, for example
or . The plugin
normalizes both forms to a base URL ending in , such as
. Leave the API key empty when the server does
not require one.

Predefined models accept audio files up to 25 MB.

Supported Models

sensevoice - Mandarin, Cantonese, English, Japanese, and Korean ASR with
emotion and audio-event tags (default)
paraformer - Chinese ASR with punctuation
paraformer-en - English ASR
fun-asr-nano - Chinese, English, Japanese, and Chinese dialect/accent ASR
with an encoder + LLM architecture