Fish Speech V1.5 - 昇腾NPU适配

NPU:                  Ascend 910B2 (9362) x2
CANN:                 8.5.1
PyTorch:              2.9.0
torch_npu:            2.9.0.post1
Python:               3.11
OS:                   Linux (aarch64)

# Install dependencies
pip install torch torch_npu
pip install modelscope loguru tiktoken soundfile numpy
pip install dac  # For vocoder support

# Set environment variables
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True

# Download the model
pip install modelscope
modelscope download --model fishaudio/fish-speech-1.5 --local_dir ./fish-speech-1.5

# Clone the fish-speech repo for tokenizer support
git clone --depth 1 https://github.com/fishaudio/fish-speech.git /tmp/fish-speech

# Basic TTS inference
python inference.py \
    --model-path ./fish-speech-1.5 \
    --text "Hello world, this is a test of Fish Speech on Ascend NPU." \
    --output output.wav

# Chinese text
python inference.py \
    --model-path ./fish-speech-1.5 \
    --text "你好，世界！这是Fish Speech在昇腾NPU上的测试。" \
    --output output_cn.wav \
    --temperature 0.7 \
    --top-p 0.9

# With custom sampling parameters
python inference.py \
    --model-path ./fish-speech-1.5 \
    --text "Your text here." \
    --output output.wav \
    --temperature 0.8 \
    --top-p 0.8 \
    --top-k 40 \
    --max-new-tokens 256

python evaluation/eval_precision.py \
    --model-path ./fish-speech-1.5 \
    --output-dir ./evaluation

python evaluation/eval_performance.py \
    --model-path ./fish-speech-1.5 \
    --num-decode-tokens 100 \
    --decode-runs 5 \
    --output-dir ./evaluation

fish-speech-1.5/
├── config.json                    # Model configuration
├── model.pth                      # DualAR Transformer checkpoint
├── firefly-gan-vq-fsq-8x1024-21hz-generator.pth  # Vocoder checkpoint
├── tokenizer.tiktoken             # Tokenizer
├── special_tokens.json            # Special tokens
├── inference.py                   # NPU inference script
├── README.md                      # This document
└── evaluation/
    ├── eval_precision.py          # Precision evaluation script
    ├── eval_performance.py        # Performance benchmark script
    ├── precision_results.json     # Precision test results
    └── performance_results.json   # Performance benchmark results

@misc{fish-speech-v1.4,
      title={Fish-Speech: Leveraging Large Language Models for Advanced Multilingual Text-to-Speech Synthesis},
      author={Shijia Liao and Yuxuan Wang and Tianyu Li and Yifan Cheng and Ruoyi Zhang and Rongzhi Zhou and Yijin Xing},
      year={2024},
      eprint={2411.01156},
      archivePrefix={arXiv},
      primaryClass={cs.SD},
      url={https://arxiv.org/abs/2411.01156},
}