Whisper Tiny ONNX - 昇腾NPU适配

# Install ModelScope
pip install modelscope

# Download model
modelscope download --model manyeyes/whisper-tiny-onnx --local_dir ./model

# Install dependencies
pip install torchaudio soundfile scipy librosa

# Generate test audio and run CPU inference (baseline)
python inference.py --audio /tmp/test_audio.wav --generate-audio --device cpu

# NPU inference
python inference.py --audio /tmp/test_audio.wav --device npu:0

# Accuracy comparison (CPU vs NPU)
python inference.py --audio /tmp/test_audio.wav --device npu:0 --compare

# Performance benchmark with 20 iterations
python inference.py --audio /tmp/test_audio.wav --device npu:0 --benchmark 20

from inference import load_model, load_audio, compute_mel, greedy_decode, decode_tokens

# Load model on NPU
model = load_model("model/tiny.pt", device="npu:0")

# Load and preprocess audio
audio = load_audio("audio.wav")
mel = compute_mel(audio, device="npu:0")

# Run inference
tokens, enc_time = greedy_decode(model, mel, language="en", device="npu:0")
text = decode_tokens(tokens)
print(text)

whisper-tiny-onnx-npu/
├── README.md              # This document
├── inference.py           # NPU inference script
├── eval_benchmark.py      # Evaluation suite
├── eval_results.json      # Evaluation results (structured)
├── run_log.log            # Full run log
└── model/                 # Model files
    ├── tiny.pt            # PyTorch checkpoint
    ├── encoder.onnx       # ONNX encoder (FP32)
    ├── encoder.int8.onnx  # ONNX encoder (INT8)
    ├── decoder.onnx       # ONNX decoder (FP32)
    ├── decoder.int8.onnx  # ONNX decoder (INT8)
    ├── conf.json          # Model config
    └── configuration.json # Additional config

@misc{whisper-tiny-onnx-npu,
  title={Whisper Tiny ONNX - Ascend NPU Adaptation},
  author={Model Agent},
  year={2026},
  note={Adapted from manyeyes/whisper-tiny-onnx on ModelScope. Evaluated on Huawei Ascend 910 NPU.}
}