fun-asr-batch

FunASR-Nano batch transcription with multi-GPU support, managed with uv.

Features

• Multi-GPU data parallelism (different files on different GPUs)
• Handles long audio by chunking into configurable segments (default 50s)
• Processes all audio files in input directory and outputs TXT files
• Supports WAV, MP3, FLAC, M4A formats (recursive directory scan)
• Skip already-transcribed files

Requirements

• Python ≥ 3.10
• CUDA-capable GPU(s)
• uv
• FunASR source (provides the model module):

git clone https://github.com/FunAudioLLM/Fun-ASR.git

Installation

# Install uv if you haven't already
curl -LsSf https://astral.sh/uv/install.sh | sh

# Install the project
uv sync

Usage

# Single GPU
uv run fun-asr-batch \
    --input_dir /path/to/audio \
    --output_dir /path/to/output \
    --model_dir FunAudioLLM/Fun-ASR-Nano-2512

# Multi-GPU (GPUs 0 and 1)
uv run fun-asr-batch \
    --input_dir /path/to/audio \
    --output_dir /path/to/output \
    --gpus 0,1

# With verbose output
uv run fun-asr-batch \
    --input_dir /path/to/audio \
    --output_dir /path/to/output \
    --verbose

Options

Option	Default	Description
--model_dir	FunAudioLLM/Fun-ASR-Nano-2512	Path to model or HuggingFace model ID
--input_dir	(required)	Directory containing audio files
--output_dir	(required)	Directory to save transcription TXT files
--max_length	3500	Maximum tokens to generate per chunk
--chunk_seconds	50	Chunk duration in seconds for long audio
--language	auto	Target language: auto, zh, en, ja
--no_itn		Disable inverse text normalization
--hotwords		Comma-separated hotwords for better recognition
--gpus	0	Comma-separated GPU IDs (e.g. 0,1,2,3)
--no_skip_existing		Re-transcribe files even if output TXT exists
--verbose		Print transcription results in real-time