FireRedASR2-AED Ascend NPU 适配部署指南

# 安装基础依赖
pip install torch torch_npu
pip install kaldiio kaldi_native_fbank sentencepiece

# 安装 ModelScope（用于下载模型）
pip install modelscope

# 下载模型
modelscope download --model xukaituo/FireRedASR2-AED --local_dir ./pretrained_models/FireRedASR2-AED

# 克隆官方代码（模型定义依赖）
git clone https://github.com/FireRedTeam/FireRedASR2S.git

git clone https://atomgit.com/weixin_62994174/FireRedASR2-AED.git
cd FireRedASR2-AED

# 使用 ffmpeg 转换音频格式
ffmpeg -i input.mp3 -ar 16000 -ac 1 -acodec pcm_s16le -f wav output.wav

import sys
sys.path.insert(0, "path/to/FireRedASR2S")
from inference import InferenceConfig, NPUASRInference

config = InferenceConfig(
    model_dir="./pretrained_models/FireRedASR2-AED",
    device="npu",        # 使用昇腾 NPU
    use_half=False,      # FP32 精度（可设为 True 使用 FP16）
    beam_size=3,
)

asr = NPUASRInference(config)
asr.load_model()

results = asr.transcribe(["sample.wav"])
for r in results:
    print(f"识别结果: {r['text']}")
    print(f"置信度: {r['confidence']}")
    print(f"RTF: {r['rtf']}")

python3 inference.py \
  --model_dir ./pretrained_models/FireRedASR2-AED \
  --device npu \
  --wav_path sample1.wav sample2.wav \
  --beam_size 3 \
  --output result.json

# NPU vs CPU 一致性验证
python3 evaluate_accuracy.py \
  --model_dir ./pretrained_models/FireRedASR2-AED \
  --output accuracy_results.json

# 带标注数据的完整评估
python3 evaluate_accuracy.py \
  --model_dir ./pretrained_models/FireRedASR2-AED \
  --manifest test_manifest.txt \
  --output accuracy_results.json

# NPU 性能基准 + CPU 对比
python3 evaluate_performance.py \
  --model_dir ./pretrained_models/FireRedASR2-AED \
  --num_test_files 10 \
  --audio_duration 5.0 \
  --batch_size 1 \
  --warmup 3 \
  --runs 10 \
  --compare_cpu \
  --output performance_results.json

# 仅 NPU 性能基准
python3 evaluate_performance.py \
  --model_dir ./pretrained_models/FireRedASR2-AED \
  --num_test_files 5 \
  --audio_duration 3.0 \
  --batch_size 1 \
  --output performance_results.json

FireRedASR2-AED/
├── inference.py                # NPU 推理主脚本
├── evaluate_accuracy.py         # 精度评估脚本（NPU vs CPU）
├── evaluate_performance.py      # 性能基准测试脚本
├── accuracy_results.json        # 精度评估结果
├── performance_results.json     # 性能测试结果
├── inference_output.json        # 推理输出样例
├── README.md                    # 本文档
├── test_tone.wav                # 测试音频（合成）
└── FireRedASR2S/               # 官方模型代码（需自行克隆）

2026-05-16 05:30:07,467 INFO: NPU available: True, count: 2
2026-05-16 05:30:07,472 INFO: NPU device: Ascend910_9362
2026-05-16 05:30:07,472 INFO: NPU count: 2
2026-05-16 05:30:07,472 INFO: Using device: npu
2026-05-16 05:30:07,473 INFO: Loading feature extractor...
2026-05-16 05:30:07,473 INFO: Loading AED model...
2026-05-16 05:30:17,086 INFO: #param of FireRedAsrAed is 1152521275 = 4396.5 MB (float32)
2026-05-16 05:30:20,254 INFO: Model loaded successfully on npu
============================================================
Utterance: test_tone
Text:      
Confidence: 0.914
Duration:   2.0s
RTF:        0.2174

2026-05-16 05:31:25,188 INFO: ACCURACY EVALUATION SUMMARY
2026-05-16 05:31:25,188 INFO: ============================================================
2026-05-16 05:31:25,188 INFO:   Samples:         1
2026-05-16 05:31:25,188 INFO:   Output Match:    100.0%
2026-05-16 05:31:25,188 INFO:   CER (NPU vs CPU): Avg=0.0%, Max=0.0%
2026-05-16 05:31:25,188 INFO:   Accuracy Gate:    PASS

2026-05-16 05:33:36,230 INFO: PERFORMANCE BENCHMARK SUMMARY
2026-05-16 05:33:36,230 INFO: ============================================================
2026-05-16 05:33:36,230 INFO:   Throughput:     32.2937x realtime
2026-05-16 05:33:36,230 INFO:   Avg Latency:    0.0929s
2026-05-16 05:33:36,230 INFO:   Avg RTF:        0.029124
2026-05-16 05:33:36,230 INFO:   P50/P90/P99:    0.093s / 0.0937s / 0.0939s
2026-05-16 05:33:36,230 INFO:   NPU Memory:     4582.16MB allocated
2026-05-16 05:33:36,230 INFO:   NPU vs CPU:     48.83x speedup