xujiashuai/timm-speech_eres2net_large_five_lre_8k
Model Introduction
Files & versions
Pull Requests
Discussion
Analysis

speech_eres2net_large_five_lre_8k - Ascend NPU 适配

1. 模型简介

基于 ERes2Net (Enhanced Res2Net) 大模型架构的五语言识别模型,支持粤语、英语、日语、韩语、中文五种语言的识别,输入为 8kHz 采样率音频。

  • 原始模型: iic/speech_eres2net_large_five_lre_8k
  • 框架: PyTorch
  • 任务: Language Recognition (5 languages)
  • Backbone: ERes2Net-Large (m_channels=64, embed_dim=512)

2. 昇腾 NPU 适配结果

指标值
Embedding Cosine Similarity1.000000
Classifier Cosine Similarity1.000000
平均延迟13.84ms
输出维度512
推理精度float32
设备Ascend 910B4

3. 环境要求

组件版本
CANN8.5.1
torch_npu2.9.0.post1
PyTorch2.9.0
Python3.11
modelscopelatest

4. 快速使用

# 设置环境
export ASCEND_HOME_PATH=/usr/local/Ascend/cann
export LD_LIBRARY_PATH=/usr/local/Ascend/cann/lib64:$LD_LIBRARY_PATH

# 运行推理
python inference.py

# 运行精度验证
python evaluate.py

5. 推理输出证据

NPU 推理输出 (float32, 8kHz audio):

Model: iic/speech_eres2net_large_five_lre_8k
Device: npu:0
Embedding shape: [1, 512]
Embedding sample: [-0.0015733233885839581, -0.0009446564363315701, -0.0008025005809031427, -0.0073548611253499985, -0.009050089865922928]
Predicted language: English (confidence: 0.9971)
Latency: 13.84ms

--- CPU vs NPU 精度对比 ---
Embedding Cosine Similarity: 1.000000
Classifier Cosine Similarity: 1.000000
Max Abs Error: 0.000442
CPU prediction: English
NPU prediction: English
Predictions match: True
SUCCESS

6. CPU vs NPU 精度对比

指标CPU (float32)NPU (float32)误差
Embedding Cosine Similarity基准1.000000< 0.000001
Classifier Cosine Similarity基准1.000000< 0.000001
Max Abs Error-0.000442-
输出维度512512一致
预测语言EnglishEnglish一致

7. 模型结构

  • Backbone: ERes2Net-Large with AFF (Attentional Feature Fusion)
  • 输入: 80-dim fbank 特征 (8kHz)
  • Pooling: TSTP (Temporal Statistics Pooling)
  • Embedding: 512 维
  • Classifier: Linear(512, 5) → 5 languages
  • Languages: Cantonese, English, Japanese, Korean, Chinese

8. 验证报告

详见 report.json。