Fun-ASR-MLT-Nano-2512

# Install dependencies
pip install funasr soundfile librosa

# Download model from ModelScope
pip install modelscope
modelscope download --model FunAudioLLM/Fun-ASR-MLT-Nano-2512 --local_dir ./Fun-ASR-MLT-Nano-2512

import torch
import torch_npu

# Ensure NPU is available
assert torch.npu.is_available(), "NPU not available"
device = torch.device("npu:0")

# Load and run
from inference import load_model_cpu, move_model_to_npu, run_inference_npu

model, kwargs = load_model_cpu("./Fun-ASR-MLT-Nano-2512")
model_npu = move_model_to_npu(model, device)

result = run_inference_npu(
    model_npu, kwargs,
    wav_path="./example/zh.mp3",
    language="中文",
    itn=True,
)
print(result["text"])

python inference.py --wav example/zh.mp3 --language "中文"
python inference.py --all_examples  # run all examples

Fun-ASR-MLT-Nano-2512/
├── inference.py          # NPU-adapted inference script
├── evaluate.py           # Accuracy & performance benchmark script
├── README.md             # This file
├── model.py              # Model architecture definition
├── model.pt              # Model checkpoint (1.97GB)
├── config.yaml           # Model configuration
├── configuration.json    # Pipeline configuration
├── ctc.py                # CTC decoder module
├── multilingual.tiktoken # CTC tokenizer vocabulary
├── tools/
│   └── utils.py          # Forced alignment utilities
├── Qwen3-0.6B/           # Qwen LLM tokenizer files
│   └── tokenizer.json, vocab.json, merges.txt, config.json
└── example/              # Test audio files (5 languages)
    ├── zh.mp3
    ├── en.mp3
    ├── ja.mp3
    ├── ko.mp3
    └── yue.mp3

Audio Input (16kHz)
    │
    ▼
WavFrontend (80-dim mel fbank, lfr_m=7, lfr_n=6)
    │
    ▼
SenseVoiceEncoderSmall (512-dim, 50 SANM blocks, 4 heads)
    │
    ├──► CTC Decoder (5-layer Transformer) ──► CTC Alignments
    │
    ▼
Audio Adaptor (2-layer Transformer, 512→1024)
    │
    ▼
Qwen3-0.6B LLM (frozen, bfloat16)
    │
    ▼
Text Output (31 languages)