nemotron-speech-streaming-en-0.6b Ascend NPU 部署指南

/data/ysws/agentsp/5-14/nemotron-speech-streaming-en-0.6b-ascend/
├── inference.py          # 精度测试脚本
├── log.txt               # 测试日志
├── README.md             # 本文档
├── test_audio_0.wav      # 测试音频样本
├── test_audio_1.wav
├── test_audio_2.wav
└── fusion_result.json    # 融合结果

docker exec -it test-modelagent bash

source /usr/local/Ascend/ascend-toolkit/set_env.sh

pip install nemo-toolkit nemo_toolkit_asr -i https://repo.huaweicloud.com/repository/pypi/simple/ --trusted-host repo.huaweicloud.com

cd /data/ysws/agentsp/5-14/nemotron-speech-streaming-en-0.6b-ascend/

# 使用默认测试音频 (随机噪声)
python3 inference.py

# 使用指定音频文件
python3 inference.py --audio_path /path/to/your/audio.wav

# 指定设备
python3 inference.py --device npu:0

cd /data/ysws/agentsp/5-14/nemotron-speech-streaming-en-0.6b-ascend/

# 运行完整精度测试
python3 inference.py --precision_test

# 指定测试张量数量
python3 inference.py --precision_test --num_tensors 20

# 组合使用
python3 inference.py --audio_path /path/to/audio.wav --precision_test

import torch
import numpy as np
from nemo.collections.asr.models import EncDecRNNTBPEModel

model_path = "/data/ysws/agentsp/5-14/nemotron-speech-streaming-en-0.6b/nemotron-speech-streaming-en-0.6b.nemo"

model = EncDecRNNTBPEModel.restore_from(model_path)
model = model.to("npu:0")
model.eval()

audio = np.random.randn(16000).astype(np.float32) * 0.01
audio_tensor = torch.from_numpy(audio).float().to("npu:0")

with torch.no_grad():
    hypothesis = model.transcribe(audio_tensor)[0]

transcription = hypothesis.text
print(f"Transcription: {transcription}")

import scipy.io.wavfile as wavfile

sample_rate, audio_data = wavfile.read("your_audio.wav")
audio_data = audio_data.astype(np.float32) / 32768.0

audio_tensor = torch.from_numpy(audio_data).float().to("npu:0")

with torch.no_grad():
    hypothesis = model.transcribe(audio_tensor)[0]

print(f"Transcription: {hypothesis.text}")