Fish Audio S2 Pro Ascend NPU 适配部署指南

# 验证 NPU 可用性
python -c "import torch; import torch_npu; print(torch.npu.device_count())"
# 输出: 8

# 验证 NPU 型号
python -c "import torch; print(torch.npu.get_device_name(0))"
# 输出: Ascend910B2C

cd fish-audio-s2-pro-ascend
pip install -r requirements.txt

# 从 ModelScope 下载
modelscope download --model fishaudio/s2-pro --local_dir ./weights

# 单卡 NPU 推理
python inference.py \
  --text "你好，欢迎使用 Fish Audio S2 Pro 昇腾适配版。[whisper]这是耳语效果演示[/whisper]" \
  --output output.wav \
  --device npu:0 \
  --dtype float16

# 输出：
#   output.wav   - 生成的音频文件 (44.1kHz, 16-bit)

{
  "accuracy_test": { "cosine_similarity": 0.9953, "relative_error_pct": 0.47, "status": "PASSED" },
  "performance_test": { "rtf_range": "0.20-0.40", "max_throughput_tps": 11476.5, "status": "PASSED" },
  "multi_language_test": { "languages": 10, "status": "PASSED" },
  "inline_control_test": { "tags": 6, "status": "PASSED" },
  "conclusion": { "overall": "PASSED" }
}

cd benchmark
python accuracy_eval.py \
  --phase npu \
  --npu-device npu:0 \
  --model-dir ../weights \
  --test-corpus ../assets/test_corpus.json \
  --output ../eval_results/accuracy_report.json

python accuracy_eval.py \
  --phase compare \
  --model-dir ../weights \
  --test-corpus ../assets/test_corpus.json \
  --output ../eval_results/accuracy_report.json

cd benchmark
python perf_eval.py \
  --device npu:0 \
  --model-dir ../weights \
  --batch-sizes 1 2 4 8 \
  --seq-len 128 \
  --num-runs 10 \
  --output ../eval_results/perf_report.json

python benchmark/generate_report.py \
  --accuracy eval_results/accuracy_report.json \
  --perf eval_results/perf_report.json \
  --output eval_results/evaluation_summary.md

fish-audio-s2-pro-ascend/
├── README.md                                    # 本文件（模型卡片 + 评测报告）
├── inference.py                                 # NPU 推理脚本（单条 / 流式）
├── requirements.txt                             # Python 依赖
├── benchmark/
│   ├── accuracy_eval.py                         # 精度评测（NPU vs CPU）
│   ├── perf_eval.py                             # 性能压测（延迟 / 吞吐 / RTF）
│   └── generate_report.py                       # 报告生成器
├── eval_results/
│   ├── accuracy_report.json                     # 精度评测原始数据
│   ├── perf_report.json                         # 性能评测原始数据
│   └── evaluation_summary.md                    # 评测汇总
├── eval_materials/
│   ├── comprehensive_evaluation_report.json     # 综合评估报告
│   └── inference_run.log                        # 推理运行日志
├── assets/
│   └── test_corpus.json                         # 评测用例（多语言 / 多标签）
└── weights/                                     # 模型权重（需下载）

@misc{liao2026fishaudios2technical,
      title={Fish Audio S2 Technical Report},
      author={Shijia Liao and Yuxuan Wang and Songting Liu and Yifan Cheng
              and Ruoyi Zhang and Tianyu Li and Shidong Li and Yisheng Zheng
              and Xingwei Liu and Qingzheng Wang and Zhizhuo Zhou and Jiahua Liu
              and Xin Chen and Dawei Han},
      year={2026},
      eprint={2603.08823},
      archivePrefix={arXiv},
      primaryClass={cs.SD},
      url={https://arxiv.org/abs/2603.08823},
}