Pocket TTS - Ascend NPU 适配版

pip install pocket-tts safetensors scipy numpy beartype

pip install modelscope
modelscope download --model AI-ModelScope/pocket-tts-without-voice-cloning

from inference import load_model_on_npu, create_voice_state, run_inference
import scipy.io.wavfile

# Load model on NPU
model = load_model_on_npu()

# Create voice state
voice_state = create_voice_state(model, seed=42)

# Generate speech
audio, stats = run_inference(model, voice_state, "Hello world!", seed=42)

# Save audio
scipy.io.wavfile.write("output.wav", model.sample_rate, audio.cpu().numpy())
print(f"Generated {stats['duration_s']:.1f}s audio in {stats['elapsed_s']:.2f}s ({stats['rtf']:.1f}x RTF)")

python inference.py --text "Hello world, this is a test." --output output.wav
python inference.py --text "Hello world." --compare   # accuracy validation
python inference.py --benchmark                        # performance test

pocket-tts-npu/
├── inference.py           # Main NPU inference script
├── README.md              # This file
├── accuracy_report.json   # Accuracy validation results
├── benchmark_report.json  # Performance benchmark results
└── output_npu.wav         # Sample generated audio

@article{orsini2025pocket,
  title={Pocket TTS: Lightweight Text-to-Speech},
  author={Orsini, Manu and Rouard, Simon and De Marmiesse, Gabriel and Volhejn, Václav and Zeghidour, Neil and Défossez, Alexandre},
  journal={arXiv preprint arXiv:2509.06926},
  year={2025}
}