AnzezxianxHazardInspectAI
/
SpeechRecognition
2
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

代码优化 文件精简

This commit is contained in:
yueliuli 2026-04-30 17:16:00 +08:00
parent 442afff195
commit 1e9906ef1a
20 changed files with 389 additions and 2133 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

280
BATCH_USAGE.md
View File

@ -1,280 +0,0 @@
# 并发批量处理使用指南
## 功能特性
✅ **运行时自动清空 temp 目录**
✅ **并发批处理** - 根据 GPU 显存/CPU 核心数自动调整并发数
✅ **预提取 WAV** - 每个视频在处理前提取音频到 temp
✅ **结果合并** - 使用 map_speaker 合并 ASR 和说话人分离结果
✅ **独立输出** - 每个视频结果分别存入 output 目录
## 使用方法
### 1. 配置视频文件夹
编辑 `main.py` 中的 `VIDEO_FOLDER` 变量：
```python
# 视频文件夹路径（全局变量）
VIDEO_FOLDER = r"D:\Userfile\Projects\AnzezxianxHazardInspectAI\Code\audio"
```
**程序会自动：**
- ✅ 扫描文件夹中的所有视频文件
- ✅ 支持格式：mp4, avi, mkv, mov, flv, wmv, m4v
- ✅ 按文件名自动排序（时间戳格式的文件名会按时间顺序排列）
**文件名格式示例：**
```
VID_20251031_132320_019.mp4 → 2025-10-31 13:23:20
VID_20251031_140530_020.mp4 → 2025-10-31 14:05:30
VID_20251101_090000_021.mp4 → 2025-11-01 09:00:00
```
### 2. 运行批处理
```bash
# 激活虚拟环境
funasr_env\Scripts\activate
# 运行批处理
python main.py
```
## 工作流程
```
开始
↓
清空 temp/ 目录
↓
创建 output/ 目录
↓
并发处理每个视频:
1. 提取 WAV 到 temp/
2. 加载 ASR 模型
3. 执行语音识别
4. 加载说话人分离模型
5. 执行说话人分离
6. 合并结果（map_speaker）
7. 保存结果到 output/
8. 清理临时 WAV
↓
生成汇总报告 output/batch_summary.json
↓
清空 temp/ 目录
↓
完成
```
## 输出文件
### 单个视频结果
`output/{video_name}_result.json`
```json
{
"total_sentences": 50,
"sentences": [
{
"speaker": "speaker_0",
"text": "你好，请问这里是哪里？",
"begin_time": 0.50,
"end_time": 2.30,
"duration": 1.80
}
]
}
```
### 汇总报告
`output/batch_summary.json`
```json
{
"total_videos": 3,
"success_count": 3,
"failed_count": 0,
"total_time_seconds": 245.67,
"results": [
{
"video": "VID_20251031_132320_019.mp4",
"success": true,
"output": "output/VID_20251031_132320_019_result.json",
"total_sentences": 50,
"speaker_counts": {
"speaker_0": 25,
"speaker_1": 25
}
}
]
}
```
## 并发策略
### GPU 模式
- 根据显存自动调整并发数
- 每个视频约需 2-3GB 显存
- 公式：`并发数 = max(1, 显存总量 / 3GB)`
### CPU 模式
- 使用 CPU 核心数作为并发数
- 使用 `multiprocessing.cpu_count()` 获取
## 性能优化建议
### 1. GPU 用户
- 确保安装 CUDA 版本 PyTorch
- 8GB 显存：建议并发 2-3
- 12GB 显存：建议并发 4
- 24GB 显存：建议并发 8
### 2. CPU 用户
- 减少并发数避免内存不足
- 建议：`并发数 = CPU 核心数 / 2`
### 3. 内存优化
每个进程约需：
- ASR 模型：2-3GB
- 说话人分离模型：1-2GB
- 总计：3-5GB/进程
确保系统内存充足：`并发数 × 5GB < 可用内存`
## 自定义配置
### 调整并发数
编辑 `main.py` 的 `main()` 函数：
```python
# 固定并发数为 2
results = process_batch_concurrent(video_paths, max_workers=2)
```
### 修改说话人分离参数
编辑 `process_single_video()` 函数：
```python
diar_service = DiarizationService(
embedding_model="eres2netv2", # campplus/eres2net/eres2netv2
device="auto",
cluster_threshold=0.5, # 0.0-1.0，越高越严格
min_cluster_size=10 # 每个说话人最少片段数
)
```
### 修改 ASR 模型
编辑 `process_single_video()` 函数：
```python
asr_service = ASRService(
model_name="paraformer-zh", # 或 "SenseVoice"
device="auto"
)
```
## 常见问题
### Q: 如何添加更多视频？
**A:** 只需将视频文件放入 `VIDEO_FOLDER` 指定的文件夹即可，程序会自动扫描。
### Q: 如何跳过某些视频？
**A:** 将这些视频移到其他文件夹，或修改 `SUPPORTED_VIDEO_FORMATS` 排除特定格式。
### Q: 处理中断了怎么办？
**A:** 重新运行即可，会自动清空 temp 目录，已完成的视频不会重复处理。
### Q: 如何查看处理进度？
**A:** 控制台会实时显示：
- 每个视频的处理状态
- 进度百分比
- 预计剩余时间
- 最终汇总报告
## 目录结构
```
audio2/
├── main.py # 主程序
├── asr_service.py # ASR 服务
├── diarization_service.py # 说话人分离服务
├── map_speaker.py # 结果合并逻辑
├── temp/ # 临时目录（运行时清空）
└── output/ # 输出目录
├── video1_result.json
├── video2_result.json
└── batch_summary.json
```
## 依赖要求
- Python 3.10+
- FunASR 1.3+
- PyTorch 2.0+
- ffmpeg（用于提取音频）
- 3D-Speaker（说话人分离）
## 运行示例
```
============================================================
并发批量语音识别处理系统
============================================================
============================================================
清空临时目录...
============================================================
✓ 已删除：D:\...\audio2\temp
✓ 已创建：D:\...\audio2\temp
✓ 输出目录：D:\...\audio2\output
找到 1 个视频文件
- VID_20251031_132320_019.mp4
============================================================
并发批处理配置
============================================================
视频数量：1
最大并发：2
CPU 核心数：8
GPU: NVIDIA GeForce RTX 3060
[VID_20251031_132320_019.mp4] 加载 ASR 模型...
[VID_20251031_132320_019.mp4] 执行语音识别...
[VID_20251031_132320_019.mp4] 加载说话人分离模型...
[VID_20251031_132320_019.mp4] 执行说话人分离...
[VID_20251031_132320_019.mp4] 合并结果...
[VID_20251031_132320_019.mp4] ✓ 处理完成
- 句子数：50
- 说话人：{'speaker_0': 25, 'speaker_1': 25}
============================================================
处理完成汇总
============================================================
总耗时：123.4s
平均每个视频：123.4s
成功：1/1
失败：0
汇总报告：output\batch_summary.json
============================================================
清理临时文件...
============================================================
清空临时目录...
============================================================
✓ 全部完成！
输出目录：D:\...\audio2\output
```

107
CLEANUP_SUMMARY.md Normal file
View File

@ -0,0 +1,107 @@
# 代码清理总结
## 清理目标
以 `main.py` 的流程为主，删除其他文件中的未使用代码。
## 当前流程（main.py）
```
阶段 1: 说话人分离 (3D-Speaker)
↓
清理 CUDA 缓存
↓
阶段 2: ASR 识别 + 合并结果
```
## 清理内容
### 1. asr_service.py
**删除的功能**:
- ❌ `use_3d_speaker` 参数及相关逻辑（已在 main.py 中手动处理）
- ❌ `_merge_diarization_segments()` 方法（未使用）
- ❌ `_map_asr_to_speaker()` 方法（已在 main.py 中内联实现）
**保留的功能**:
- ✅ `recognize()` - 基础 ASR 识别
- ✅ `_parse_result()` - 解析识别结果
- ✅ `export_to_json()` / `export_to_srt()` - 导出功能
**修改说明**:
- ASR 识别结果中的默认说话人统一设为 `speaker_0`
- 不再在 ASR 服务内部调用 3D-Speaker
### 2. map_speaker.py
**删除的功能**:
- ❌ `find_speaker()` 函数（已在 main.py 中内联实现）
- ❌ `main()` 函数（过时的示例代码）
**保留的功能**:
- ✅ `load_json()` - 加载 JSON 文件
- ✅ `save_json()` - 保存 JSON 文件
### 3. example_usage.py
**修改内容**:
- 更新输出示例中的说话人格式：`SPEAKER_00` → `speaker_0`
- 保持示例代码的参考价值
### 4. 删除的文件
- ❌ `test_staged.py` - 临时测试脚本
- ❌ `test_model_load.py` - 临时测试脚本
## 核心逻辑（main.py）
### 阶段 1: 说话人分离
```python
diar_service = DiarizationService()
diar_service._load_model()
for video in videos:
wav = extract_wav(video)
segments = diar_service.diarize(wav)
save_json(temp_file, {"segments": segments})
```
### 阶段 2: ASR 识别 + 合并
```python
asr_service = ASRService()
asr_service._load_model()
for video in videos:
asr_sentences = asr_service.recognize(wav)
# 合并说话人（只使用 3D-Speaker 结果）
for sentence in asr_sentences:
matched_speaker = 查找最大重叠的说话人
if matched_speaker:
sentence.speaker = matched_speaker
else:
sentence.speaker = "speaker_0"
export_to_json(output_file, asr_sentences)
```
## 优势
1. **逻辑清晰**: 只在一个地方（main.py）处理说话人合并
2. **避免重复**: 删除了多处重复的说话人对齐逻辑
3. **易于维护**: 核心流程集中在 main.py，服务类只负责基础功能
4. **统一格式**: 所有说话人标签统一为 `speaker_0`, `speaker_1`, ...
## 文件依赖关系
```
main.py
├── asr_service.py (基础 ASR 识别)
├── diarization_service.py (说话人分离)
└── map_speaker.py (JSON 工具函数)
```
## 未清理的文件
- `server.py` - Web API 服务（独立功能）
- `test_asr.py` - 测试脚本（可保留）
- `example_usage.py` - 示例代码（已更新）

196
INSTALL.md Normal file
View File

@ -0,0 +1,196 @@
# 项目依赖安装指南
## 📋 环境要求
- **Python**: 3.10+
- **CUDA**: 11.8+ (可选，用于 GPU 加速)
- **系统**: Windows 10/11, Linux, macOS
## 🚀 快速安装
### 1. 创建虚拟环境
```bash
python -m venv funasr_env
```
### 2. 激活虚拟环境
**Windows:**
```bash
funasr_env\Scripts\activate
```
**Linux/macOS:**
```bash
source funasr_env/bin/activate
```
### 3. 安装 PyTorch (带 CUDA 支持)
**CUDA 11.8:**
```bash
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
```
**CPU 版本:**
```bash
pip install torch torchvision torchaudio
```
### 4. 安装 3D-Speaker
```bash
# 克隆 3D-Speaker 项目到父目录
cd ..
git clone https://github.com/alibaba-damo-academy/3D-Speaker.git
# 安装 3D-Speaker 依赖
cd 3D-Speaker
pip install -e .
```
### 5. 安装其他依赖
```bash
# 返回项目目录
cd ../audio2
# 安装 requirements.txt
pip install -r requirements.txt
```
## 📦 依赖说明
### 核心依赖
| 包名 | 用途 | 必需 |
|------|------|------|
| torch | 深度学习框架 | ✅ |
| funasr | 语音识别引擎 | ✅ |
| modelscope | 模型下载与管理 | ✅ |
| speakerlab | 3D-Speaker 说话人分离 | ✅ |
| soundfile | 音频文件读写 | ✅ |
| librosa | 音频分析 | ✅ |
### 可选依赖
| 包名 | 用途 | 何时需要 |
|------|------|----------|
| onnxruntime-gpu | ONNX 推理加速 | 需要更高性能时 |
| Flask | Web API 服务 | 需要部署 Web 服务时 |
| SQLAlchemy | 数据库 ORM | 需要持久化存储时 |
## 🔧 验证安装
运行测试脚本验证安装：
```bash
# 测试模型加载
python test_model_load.py
# 运行主程序
python main.py
```
## ⚠️ 常见问题
### 1. CUDA 版本不匹配
**错误信息:**
```
RuntimeError: CUDA error: no kernel image is available for execution
```
**解决方案:**
```bash
# 卸载当前 PyTorch
pip uninstall torch torchvision torchaudio
# 根据 CUDA 版本重新安装
# CUDA 11.8
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
# CUDA 12.1
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
```
### 2. 3D-Speaker 导入失败
**错误信息:**
```
ModuleNotFoundError: No module named 'speakerlab'
```
**解决方案:**
```bash
# 确保 3D-Speaker 在项目父目录
# 结构应为:
# project/
# ├── audio2/
# └── 3D-Speaker/
# 重新安装 3D-Speaker
cd 3D-Speaker
pip install -e .
```
### 3. 模型下载失败
**错误信息:**
```
ConnectionError: Failed to download model from ModelScope
```
**解决方案:**
```bash
# 使用阿里云镜像
export MODELSCOPE_CACHE="./models"
# 或手动下载模型后放入缓存目录
```
### 4. 内存不足
**错误信息:**
```
RuntimeError: CUDA out of memory
```
**解决方案:**
- 减少并发数：修改 `main.py` 中的 `max_workers=1`
- 使用 CPU 模式：`device='cpu'`
- 关闭其他占用 GPU 的程序
## 📝 依赖版本锁定
如需精确控制版本，使用：
```bash
# 生成当前环境的依赖快照
pip freeze > requirements.lock.txt
# 使用锁定的版本安装
pip install -r requirements.lock.txt
```
## 🎯 最小化安装
如果只需要基础功能：
```bash
# 最小依赖集
pip install torch funasr modelscope soundfile scipy numpy tqdm pyyaml
```
## 📊 磁盘空间需求
| 组件 | 空间需求 |
|------|----------|
| 基础依赖 | ~2 GB |
| PyTorch (CUDA) | ~3 GB |
| FunASR 模型 | ~2 GB |
| 3D-Speaker 模型 | ~1 GB |
| **总计** | **~8 GB** |
建议预留 **10 GB** 以上可用空间。

106
asr_service.py
View File

@ -143,75 +143,11 @@ class ASRService:
print(f"模型加载完成！") print(f"模型加载完成！")
def _merge_diarization_segments(
self,
segments: List[Dict],
min_duration: float = 0.3,
merge_gap: float = 0.5
) -> List[Dict]:
"""合并相邻的同一说话人片段"""
if not segments:
return []
filtered = [s for s in segments if s["end_time"] - s["begin_time"] >= min_duration]
if not filtered:
return []
merged = [dict(filtered[0])]
for seg in filtered[1:]:
last = merged[-1]
if seg["speaker"] == last["speaker"] and seg["begin_time"] - last["end_time"] <= merge_gap:
last["end_time"] = seg["end_time"]
last["duration"] = last["end_time"] - last["begin_time"]
else:
merged.append(dict(seg))
return merged
def _map_asr_to_speaker(
self,
asr_segments: List[Dict],
diarization_segments: List[Dict]
) -> List[Dict]:
"""将 ASR 识别结果与说话人分离结果对齐"""
if not diarization_segments:
return asr_segments
aligned = []
for asr_seg in asr_segments:
asr_begin = asr_seg["begin_time"]
asr_end = asr_seg["end_time"]
best_speaker = "speaker_0"
best_overlap = 0.0
for dia_seg in diarization_segments:
dia_begin = dia_seg["begin_time"]
dia_end = dia_seg["end_time"]
overlap_start = max(asr_begin, dia_begin)
overlap_end = min(asr_end, dia_end)
overlap = max(0, overlap_end - overlap_start)
if overlap > best_overlap:
best_overlap = overlap
best_speaker = dia_seg["speaker"].replace("speaker_", "SPEAKER_")
asr_seg["speaker"] = best_speaker
aligned.append(asr_seg)
return aligned
def recognize( def recognize(
self, self,
audio_path: Union[str, Path], audio_path: Union[str, Path],
batch_size_s: int = 300, batch_size_s: int = 300,
return_raw: bool = False, return_raw: bool = False,
use_3d_speaker: bool = False,
embedding_model: str = "eres2netv2",
cluster_threshold: float = 0.5,
min_cluster_size: int = 10
) -> Union[List[Sentence], Dict]: ) -> Union[List[Sentence], Dict]:
""" """
识别音频文件 识别音频文件
@ -220,10 +156,6 @@ class ASRService:
audio_path: 音频文件路径 audio_path: 音频文件路径
batch_size_s: 批处理时长（秒） batch_size_s: 批处理时长（秒）
return_raw: 是否返回原始结果 return_raw: 是否返回原始结果
use_3d_speaker: 是否使用 3D-Speaker 替换说话人（结果保存前替换）
embedding_model: 3D-Speaker 说话人嵌入模型
cluster_threshold: 3D-Speaker 聚类阈值
min_cluster_size: 3D-Speaker 最小聚类大小
Returns: Returns:
List[Sentence]: 识别结果列表（默认） List[Sentence]: 识别结果列表（默认）
@ -251,38 +183,6 @@ class ASRService:
return result return result
sentences = self._parse_result(result) sentences = self._parse_result(result)
if use_3d_speaker and sentences:
print("正在使用 3D-Speaker 替换说话人信息...")
from diarization_service import DiarizationService
diar = DiarizationService(
embedding_model=embedding_model,
cluster_threshold=cluster_threshold,
min_cluster_size=min_cluster_size
)
dia_segments = diar.diarize(audio_path)
diarization_segments = [
{"speaker": s.speaker, "begin_time": s.begin_time, "end_time": s.end_time}
for s in dia_segments
]
if self.merge_segments:
diarization_segments = self._merge_diarization_segments(
diarization_segments,
min_duration=self.min_segment_duration,
merge_gap=self.merge_gap
)
asr_segments = [s.to_dict() for s in sentences]
aligned_segments = self._map_asr_to_speaker(asr_segments, diarization_segments)
for i, seg in enumerate(aligned_segments):
sentences[i].speaker = seg["speaker"]
print(f"说话人信息已替换，最终识别出 {len(sentences)} 句话")
return sentences return sentences
def _parse_result(self, result: List[Dict]) -> List[Sentence]: def _parse_result(self, result: List[Dict]) -> List[Sentence]:
@ -324,7 +224,7 @@ class ASRService:
results = [] results = []
for audio_path in audio_paths: for audio_path in audio_paths:
try: try:
result = self.recognize(audio_path, batch_size_s, use_3d_speaker=use_3d_speaker) result = self.recognize(audio_path, batch_size_s)
results.append(result) results.append(result)
except Exception as e: except Exception as e:
print(f"识别失败 [{audio_path}]: {e}") print(f"识别失败 [{audio_path}]: {e}")
@ -383,7 +283,9 @@ def recognize_audio(
) -> List[Sentence]: ) -> List[Sentence]:
"""快速识别音频文件""" """快速识别音频文件"""
service = ASRService(model_name=model_name, device=device) service = ASRService(model_name=model_name, device=device)
return service.recognize(audio_path, use_3d_speaker=use_3d_speaker) result = service.recognize(audio_path, return_raw=False)
assert isinstance(result, list)
return result
if __name__ == "__main__": if __name__ == "__main__":

11
diarization_service.py
View File

@ -125,8 +125,8 @@ class DiarizationService:
} }
try: try:
from speakerlab.bin.infer_diarization import Diarization3Dspeaker from speakerlab.bin.infer_diarization import Diarization3Dspeaker # type: ignore
print(f" - 导入 Diarization3Dspeaker 完成") print(f" - 导入 Diarization3Dspeaker 完成")
sys.stdout.flush() sys.stdout.flush()
@ -136,12 +136,12 @@ class DiarizationService:
hf_access_token=self.hf_access_token, hf_access_token=self.hf_access_token,
model_cache_dir=self.cache_dir model_cache_dir=self.cache_dir
) )
print(f" - 模型实例化完成") print(f" - 模型实例化完成")
sys.stdout.flush() sys.stdout.flush()
print(f"模型加载完成！") print(f"模型加载完成！")
sys.stdout.flush() sys.stdout.flush()
except Exception as e: except Exception as e:
print(f"\n✗ 模型加载失败：{e}") print(f"\n✗ 模型加载失败：{e}")
sys.stdout.flush() sys.stdout.flush()
@ -174,6 +174,9 @@ class DiarizationService:
print(f"正在执行说话人分离: {audio_path}") print(f"正在执行说话人分离: {audio_path}")
if self.model is None:
raise RuntimeError("模型未正确加载，无法执行说话人分离")
result = self.model( result = self.model(
wav=str(audio_path), wav=str(audio_path),
speaker_num=speaker_num speaker_num=speaker_num

46
enable_long_path.ps1
View File

@ -1,46 +0,0 @@
# 启用 Windows 长路径支持（需要管理员权限）
# 运行后重启电脑生效
Write-Host "========================================" -ForegroundColor Cyan
Write-Host "启用 Windows 长路径支持" -ForegroundColor Cyan
Write-Host "========================================" -ForegroundColor Cyan
Write-Host ""
# 检查是否以管理员身份运行
if (-NOT ([Security.Principal.WindowsPrincipal][Security.Principal.WindowsIdentity]::GetCurrent()).IsInRole([Security.Principal.WindowsBuiltInRole] "Administrator")) {
Write-Host "❌ 请以管理员身份运行 PowerShell 后再执行此脚本" -ForegroundColor Red
Write-Host " 右键点击 PowerShell -> 以管理员身份运行" -ForegroundColor Yellow
pause
exit
}
# 启用长路径支持
Write-Host "正在启用长路径支持..." -ForegroundColor Yellow
try {
Set-ItemProperty -Path "HKLM:\SYSTEM\CurrentControlSet\Control\FileSystem" -Name "LongPathsEnabled" -Value 1
Write-Host "✅ 注册表修改成功" -ForegroundColor Green
} catch {
Write-Host "❌ 修改失败: $_" -ForegroundColor Red
pause
exit
}
# 启用 Python 长路径支持
Write-Host ""
Write-Host "Python 长路径环境变量:" -ForegroundColor Yellow
$envVar = [Environment]::GetEnvironmentVariable("PYTHONLEGACYWINDOWSFSENCODING", "User")
if ($envVar -eq $null) {
[Environment]::SetEnvironmentVariable("PYTHONLEGACYWINDOWSFSENCODING", "1", "User")
Write-Host "✅ 已设置 PYTHONLEGACYWINDOWSFSENCODING=1" -ForegroundColor Green
} else {
Write-Host " 已存在: PYTHONLEGACYWINDOWSFSENCODING=$envVar" -ForegroundColor Cyan
}
Write-Host ""
Write-Host "========================================" -ForegroundColor Green
Write-Host "✅ 设置完成！" -ForegroundColor Green
Write-Host "========================================" -ForegroundColor Green
Write-Host ""
Write-Host "注意: 需要重启电脑才能完全生效" -ForegroundColor Yellow
Write-Host ""
pause

128
example_usage.py
View File

@ -1,128 +0,0 @@
"""
FunASR 使用示例
展示常见的语音识别应用场景
"""
from asr_service import ASRService, recognize_audio
def example_1_basic_recognition():
"""示例1: 基础识别"""
print("=" * 60)
print("示例1: 基础语音识别")
print("=" * 60)
# 方式1: 使用便捷函数
# results = recognize_audio("meeting.wav")
# 方式2: 使用服务类（推荐，可复用）
service = ASRService(model_name="paraformer-zh")
# results = service.recognize("meeting.wav")
print("代码:")
print(" from asr_service import recognize_audio")
print(" results = recognize_audio('meeting.wav')")
print(" for sent in results:")
print(" print(sent)")
print()
print("输出格式:")
print(" [SPEAKER_00] 大家好，今天的会议现在开始。 (0.50s - 3.20s)")
print(" [SPEAKER_01] 好的，我先汇报一下进度。 (3.50s - 6.10s)")
def example_2_batch_processing():
"""示例2: 批量处理"""
print("\n" + "=" * 60)
print("示例2: 批量处理多个音频")
print("=" * 60)
print("代码:")
print(" from pathlib import Path")
print(" from asr_service import ASRService")
print()
print(" service = ASRService()")
print(" audio_files = list(Path('./audio').glob('*.wav'))")
print(" results = service.recognize_batch(audio_files)")
print()
print(" for audio_path, sentences in zip(audio_files, results):")
print(" print(f'{audio_path}: {len(sentences)} 句话')")
def example_3_export_results():
"""示例3: 导出结果"""
print("\n" + "=" * 60)
print("示例3: 导出识别结果")
print("=" * 60)
print("代码:")
print(" service = ASRService()")
print(" sentences = service.recognize('meeting.wav')")
print()
print(" # 导出为 JSON")
print(" service.export_to_json(sentences, 'meeting.json')")
print()
print(" # 导出为 SRT 字幕")
print(" service.export_to_srt(sentences, 'meeting.srt')")
print()
print("JSON 输出示例:")
print(""" {
"total_sentences": 2,
"sentences": [
{
"speaker": "SPEAKER_00",
"text": "大家好",
"begin_time": 0.50,
"end_time": 3.20,
"duration": 2.70
}
]
}""")
def example_4_different_models():
"""示例4: 选择不同模型"""
print("\n" + "=" * 60)
print("示例4: 选择不同模型")
print("=" * 60)
print("模型选择:")
print()
print("1. paraformer-zh (默认)")
print(" - 达摩院出品，中文识别精度高")
print(" - 支持说话人分离")
print(" - 代码: ASRService(model_name='paraformer-zh')")
print()
print("2. SenseVoice")
print(" - 多语言支持（中、英、日、韩等）")
print(" - 支持情感识别")
print(" - 代码: ASRService(model_name='SenseVoice')")
def example_5_hardware_options():
"""示例5: 硬件选择"""
print("\n" + "=" * 60)
print("示例5: 选择运行设备")
print("=" * 60)
print("设备选项:")
print()
print(" # 自动选择 (推荐)")
print(" service = ASRService(device='auto')")
print()
print(" # 使用 GPU")
print(" service = ASRService(device='cuda')")
print()
print(" # 使用 CPU")
print(" service = ASRService(device='cpu')")
if __name__ == "__main__":
example_1_basic_recognition()
example_2_batch_processing()
example_3_export_results()
example_4_different_models()
example_5_hardware_options()
print("\n" + "=" * 60)
print("提示: 运行测试请使用: python test_asr.py -f your_audio.wav")
print("=" * 60)

31
fix_single_process.bat
View File

@ -1,31 +0,0 @@
@echo off
echo ============================================================
echo 修复进程崩溃问题 - 启用单进程模式
echo ============================================================
echo.
echo 正在修改 main.py...
echo.
REM 读取文件内容并替换
powershell -Command ^
"$content = Get-Content -Path 'main.py' -Raw; ^
$content = $content -replace 'MAX_WORKERS_OVERRIDE = None', 'MAX_WORKERS_OVERRIDE = 1 # 强制单进程模式'; ^
Set-Content -Path 'main.py' -Value $content -Encoding UTF8"
echo.
echo ✓ 修改完成！
echo.
echo ============================================================
echo 已启用单进程模式
echo ============================================================
echo.
echo 现在可以运行：
echo python main.py
echo.
echo 如果需要恢复多进程模式，请编辑 main.py:
echo 找到：MAX_WORKERS_OVERRIDE = 1
echo 改为：MAX_WORKERS_OVERRIDE = None
echo.
echo ============================================================
pause

13
install_3d_speaker_deps.bat
View File

@ -1,13 +0,0 @@
@echo off
echo ========================================
echo 安装 3D-Speaker 说话人分离所需依赖
echo ========================================
pip install -r requirements_3d_speaker.txt
echo.
echo ========================================
echo 安装完成！
echo 现在可以运行: python diarization_service.py --wav input/your_audio.wav --out result.json --model eres2netv2
echo ========================================
pause

9
main.py
View File

@ -265,6 +265,8 @@ def process_batch_asr(video_paths: List[Path], diar_results: Dict, max_workers:
}) })
continue continue
wav_path = None
try: try:
print(f"\n[{i}/{len(video_paths)}] 处理：{video_path.name}") print(f"\n[{i}/{len(video_paths)}] 处理：{video_path.name}")
@ -326,6 +328,11 @@ def process_batch_asr(video_paths: List[Path], diar_results: Dict, max_workers:
# 5. 保存最终结果 # 5. 保存最终结果
output_file = OUTPUT_DIR / f"{video_path.stem}_result.json" output_file = OUTPUT_DIR / f"{video_path.stem}_result.json"
# 确保 asr_sentences 是 List[Sentence] 类型
if isinstance(asr_sentences, dict):
# 如果是字典，尝试获取 sentences 键或转换为空列表
asr_sentences = asr_sentences.get("sentences", [])
asr_service.export_to_json(asr_sentences, output_file) asr_service.export_to_json(asr_sentences, output_file)
# 统计说话人 # 统计说话人
@ -359,7 +366,7 @@ def process_batch_asr(video_paths: List[Path], diar_results: Dict, max_workers:
finally: finally:
# 清理临时文件 # 清理临时文件
if wav_path.exists(): if wav_path and wav_path.exists():
try: try:
wav_path.unlink() wav_path.unlink()
except: except:

217
requirements.txt
View File

@ -1,153 +1,64 @@
addict==2.4.0 # ============================================
aiohappyeyeballs==2.6.1 # 语音识别 + 说话人分离 项目依赖
aiohttp==3.13.5 # ============================================
aiosignal==1.4.0 # 使用说明：
alembic==1.18.4 # 1. 创建虚拟环境：python -m venv funasr_env
aliyun-python-sdk-core==2.16.0 # 2. 激活虚拟环境：funasr_env\Scripts\activate
aliyun-python-sdk-kms==2.16.5 # 3. 安装依赖：pip install -r requirements.txt
annotated-doc==0.0.4 # ============================================
antlr4-python3-runtime==4.9.3
anyio==4.13.0 # ---------- 核心框架 ----------
asteroid-filterbanks==0.4.0 torch>=2.7.0
async-timeout==5.0.1 torchaudio>=2.7.0
attrs==26.1.0 torchvision>=0.22.0
audioread==3.1.0
blinker==1.9.0 # ---------- FunASR 语音识别 ----------
certifi==2026.4.22 funasr>=1.3.0
cffi==2.0.0 modelscope>=1.36.0
charset-normalizer==3.4.7 transformers>=5.7.0
click==8.3.3
colorama==0.4.6 # ---------- 3D-Speaker 说话人分离 ----------
coloredlogs==15.0.1 # 注意：3D-Speaker 需要手动克隆到项目目录
colorlog==6.10.1 # git clone https://github.com/alibaba-damo-academy/3D-Speaker.git
contourpy==1.3.2 speakerlab>=1.0.0
crcmod==1.7
cryptography==47.0.0 # ---------- 音频处理 ----------
cycler==0.12.1 soundfile>=0.12.0
datasets==4.8.5 librosa>=0.11.0
decorator==5.2.1 scipy>=1.15.0
dill==0.4.1 numpy>=2.2.0
docopt==0.6.2
editdistance==0.8.1 # ---------- 机器学习基础库 ----------
einops==0.8.2 scikit-learn>=1.7.0
exceptiongroup==1.3.1 numba>=0.65.0
fastcluster==1.3.0 pandas>=2.3.0
filelock==3.25.2
Flask==3.1.3 # ---------- 聚类算法 ----------
flatbuffers==25.12.19 hdbscan>=0.8.42
fonttools==4.62.1 umap-learn>=0.5.0
frozenlist==1.8.0 fastcluster>=1.2.0
fsspec==2026.2.0
funasr==1.3.1 # ---------- 深度学习组件 ----------
greenlet==3.5.0 pytorch-lightning>=2.6.0
h11==0.16.0 lightning>=2.6.0
hdbscan==0.8.42 pyannote.audio>=3.4.0
hf-xet==1.4.3
httpcore==1.0.9 # ---------- 数据处理 ----------
httpx==0.28.1 datasets>=4.8.0
huggingface_hub==1.12.0 pyarrow>=24.0.0
humanfriendly==10.0 sentencepiece>=0.2.1
hydra-core==1.3.2
HyperPyYAML==1.2.3 # ---------- 工具库 ----------
idna==3.13 tqdm>=4.67.0
itsdangerous==2.2.0 pyyaml>=6.0
jaconv==0.5.0 simplejson>=3.19.0
jamo==0.4.1 sortedcontainers>=2.4.0
jieba==0.42.1 addict>=2.4.0
Jinja2==3.1.6 jieba>=0.42.0
jmespath==0.10.0
joblib==1.5.3 # ---------- 可选：ONNX 加速 ----------
julius==0.2.7 # onnxruntime-gpu>=1.23.0
kaldiio==2.18.1
kiwisolver==1.5.0 # ---------- 可选：Web API 服务 ----------
lazy-loader==0.5 # Flask>=3.1.0
librosa==0.11.0 # SQLAlchemy>=2.0.0
lightning==2.6.1
lightning-utilities==0.15.3
llvmlite==0.47.0
Mako==1.3.12
markdown-it-py==4.0.0
MarkupSafe==3.0.3
matplotlib==3.10.9
mdurl==0.1.2
modelscope==1.36.3
mpmath==1.3.0
msgpack==1.1.2
multidict==6.7.1
multiprocess==0.70.19
networkx==3.4.2
numba==0.65.1
numpy==2.2.6
omegaconf==2.3.0
onnxruntime-gpu==1.23.2
opencv-python==4.13.0.92
optuna==4.8.0
oss2==2.19.1
packaging==26.2
pandas==2.3.3
pillow==12.2.0
platformdirs==4.9.6
pooch==1.9.0
primePy==1.3
propcache==0.4.1
protobuf==7.34.1
pyannote.audio==3.4.0
pyannote.core==5.0.0
pyannote.database==5.1.3
pyannote.metrics==3.2.1
pyannote.pipeline==3.0.1
pyarrow==24.0.0
pycparser==3.0
pycryptodome==3.23.0
Pygments==2.20.0
pynndescent==0.6.0
pyparsing==3.3.2
pyreadline3==3.5.4
python-dateutil==2.9.0.post0
python_speech_features==0.6
pytorch-lightning==2.6.1
pytorch-metric-learning==2.9.0
pytorch-wpe==0.0.1
pytz==2026.1.post1
PyYAML==6.0.3
regex==2026.4.4
requests==2.33.1
rich==15.0.0
ruamel.yaml==0.18.17
ruamel.yaml.clib==0.2.15
safetensors==0.7.0
scikit-learn==1.7.2
scipy==1.15.3
semver==3.0.4
sentencepiece==0.2.1
shellingham==1.5.4
simplejson==4.1.1
six==1.17.0
sortedcontainers==2.4.0
soundfile==0.13.1
soxr==1.0.0
speechbrain==1.1.0
SQLAlchemy==2.0.49
sympy==1.14.0
tabulate==0.10.0
tensorboardX==2.6.5
threadpoolctl==3.6.0
tokenizers==0.22.2
tomli==2.4.1
torch==2.7.1+cu118
torch-audiomentations==0.12.0
torch-complex==0.4.4
torch_pitch_shift==1.2.5
torchaudio==2.7.1+cu118
torchmetrics==1.9.0
torchvision==0.22.1
tqdm==4.67.3
transformers==5.7.0
typer==0.25.0
typing_extensions==4.15.0
tzdata==2026.2
umap-learn==0.5.12
urllib3==2.6.3
Werkzeug==3.1.8
xxhash==3.7.0
yarl==1.23.0

24
requirements_3d_speaker.txt
View File

@ -1,24 +0,0 @@
funasr>=1.0.0
modelscope>=1.15.0
torch>=2.0.0
torchaudio>=2.0.0
torchvision>=0.15.0
transformers>=4.30.0
numpy>=1.24.0
scipy>=1.10.0
scikit-learn>=1.0.0
soundfile>=0.12.0
kaldiio>=2.18.0
pyyaml>=6.0
tqdm>=4.65.0
numba>=0.56.0
fastcluster>=1.2.0
umap-learn>=0.5.0
datasets>=2.0.0
opencv-python>=4.7.0
python-speech-features>=0.6.0
onnxruntime-gpu>=1.15.0
pyannote.audio>=3.0.0
simplejson>=3.19.0
sortedcontainers>=2.4.0
addict>=2.4.0

972
result.json
View File

@ -1,972 +0,0 @@
{
"total_segments": 161,
"speaker_count": 4,
"segments": [
{
"speaker": "speaker_1",
"begin_time": 2.31,
"end_time": 6.76,
"duration": 4.45
},
{
"speaker": "speaker_1",
"begin_time": 13.31,
"end_time": 14.3,
"duration": 0.99
},
{
"speaker": "speaker_2",
"begin_time": 15.21,
"end_time": 17.19,
"duration": 1.98
},
{
"speaker": "speaker_1",
"begin_time": 28.7,
"end_time": 31.32,
"duration": 2.62
},
{
"speaker": "speaker_2",
"begin_time": 31.32,
"end_time": 32.64,
"duration": 1.32
},
{
"speaker": "speaker_1",
"begin_time": 34.32,
"end_time": 35.45,
"duration": 1.12
},
{
"speaker": "speaker_2",
"begin_time": 35.45,
"end_time": 36.85,
"duration": 1.41
},
{
"speaker": "speaker_1",
"begin_time": 37.37,
"end_time": 38.22,
"duration": 0.85
},
{
"speaker": "speaker_2",
"begin_time": 38.5,
"end_time": 40.32,
"duration": 1.82
},
{
"speaker": "speaker_2",
"begin_time": 40.6,
"end_time": 42.43,
"duration": 1.83
},
{
"speaker": "speaker_1",
"begin_time": 42.71,
"end_time": 43.84,
"duration": 1.12
},
{
"speaker": "speaker_2",
"begin_time": 43.84,
"end_time": 48.48,
"duration": 4.64
},
{
"speaker": "speaker_1",
"begin_time": 50.65,
"end_time": 51.72,
"duration": 1.07
},
{
"speaker": "speaker_1",
"begin_time": 52.35,
"end_time": 53.48,
"duration": 1.12
},
{
"speaker": "speaker_2",
"begin_time": 53.48,
"end_time": 54.98,
"duration": 1.5
},
{
"speaker": "speaker_1",
"begin_time": 54.98,
"end_time": 56.08,
"duration": 1.1
},
{
"speaker": "speaker_1",
"begin_time": 57.01,
"end_time": 59.92,
"duration": 2.91
},
{
"speaker": "speaker_1",
"begin_time": 60.36,
"end_time": 62.23,
"duration": 1.88
},
{
"speaker": "speaker_0",
"begin_time": 62.23,
"end_time": 62.68,
"duration": 0.45
},
{
"speaker": "speaker_1",
"begin_time": 64.0,
"end_time": 67.38,
"duration": 3.38
},
{
"speaker": "speaker_2",
"begin_time": 67.38,
"end_time": 68.88,
"duration": 1.5
},
{
"speaker": "speaker_1",
"begin_time": 68.88,
"end_time": 69.47,
"duration": 0.59
},
{
"speaker": "speaker_1",
"begin_time": 70.67,
"end_time": 80.64,
"duration": 9.97
},
{
"speaker": "speaker_1",
"begin_time": 80.92,
"end_time": 82.05,
"duration": 1.12
},
{
"speaker": "speaker_2",
"begin_time": 82.05,
"end_time": 85.81,
"duration": 3.77
},
{
"speaker": "speaker_1",
"begin_time": 86.11,
"end_time": 88.73,
"duration": 2.62
},
{
"speaker": "speaker_2",
"begin_time": 88.73,
"end_time": 89.28,
"duration": 0.55
},
{
"speaker": "speaker_2",
"begin_time": 89.73,
"end_time": 92.65,
"duration": 2.92
},
{
"speaker": "speaker_1",
"begin_time": 102.54,
"end_time": 103.55,
"duration": 1.01
},
{
"speaker": "speaker_2",
"begin_time": 103.83,
"end_time": 105.7,
"duration": 1.88
},
{
"speaker": "speaker_1",
"begin_time": 105.7,
"end_time": 106.36,
"duration": 0.66
},
{
"speaker": "speaker_1",
"begin_time": 107.99,
"end_time": 109.3,
"duration": 1.31
},
{
"speaker": "speaker_1",
"begin_time": 109.77,
"end_time": 110.64,
"duration": 0.87
},
{
"speaker": "speaker_1",
"begin_time": 111.49,
"end_time": 113.37,
"duration": 1.88
},
{
"speaker": "speaker_1",
"begin_time": 117.81,
"end_time": 122.69,
"duration": 4.88
},
{
"speaker": "speaker_2",
"begin_time": 122.69,
"end_time": 124.94,
"duration": 2.25
},
{
"speaker": "speaker_1",
"begin_time": 124.94,
"end_time": 126.44,
"duration": 1.5
},
{
"speaker": "speaker_2",
"begin_time": 126.44,
"end_time": 132.44,
"duration": 6.0
},
{
"speaker": "speaker_1",
"begin_time": 132.44,
"end_time": 133.94,
"duration": 1.5
},
{
"speaker": "speaker_2",
"begin_time": 133.94,
"end_time": 136.57,
"duration": 2.63
},
{
"speaker": "speaker_1",
"begin_time": 136.85,
"end_time": 140.22,
"duration": 3.38
},
{
"speaker": "speaker_2",
"begin_time": 140.22,
"end_time": 143.97,
"duration": 3.75
},
{
"speaker": "speaker_1",
"begin_time": 143.97,
"end_time": 144.72,
"duration": 0.75
},
{
"speaker": "speaker_2",
"begin_time": 144.72,
"end_time": 149.39,
"duration": 4.66
},
{
"speaker": "speaker_2",
"begin_time": 149.88,
"end_time": 151.76,
"duration": 1.88
},
{
"speaker": "speaker_1",
"begin_time": 152.33,
"end_time": 154.21,
"duration": 1.88
},
{
"speaker": "speaker_2",
"begin_time": 154.21,
"end_time": 157.51,
"duration": 3.3
},
{
"speaker": "speaker_2",
"begin_time": 157.79,
"end_time": 160.75,
"duration": 2.96
},
{
"speaker": "speaker_2",
"begin_time": 161.03,
"end_time": 163.78,
"duration": 2.75
},
{
"speaker": "speaker_1",
"begin_time": 166.33,
"end_time": 169.08,
"duration": 2.75
},
{
"speaker": "speaker_1",
"begin_time": 171.87,
"end_time": 173.0,
"duration": 1.12
},
{
"speaker": "speaker_2",
"begin_time": 173.0,
"end_time": 174.5,
"duration": 1.5
},
{
"speaker": "speaker_1",
"begin_time": 174.5,
"end_time": 176.59,
"duration": 2.09
},
{
"speaker": "speaker_1",
"begin_time": 177.39,
"end_time": 178.59,
"duration": 1.2
},
{
"speaker": "speaker_1",
"begin_time": 183.24,
"end_time": 186.62,
"duration": 3.38
},
{
"speaker": "speaker_2",
"begin_time": 186.62,
"end_time": 188.87,
"duration": 2.25
},
{
"speaker": "speaker_1",
"begin_time": 188.87,
"end_time": 190.37,
"duration": 1.5
},
{
"speaker": "speaker_0",
"begin_time": 190.37,
"end_time": 190.92,
"duration": 0.55
},
{
"speaker": "speaker_1",
"begin_time": 191.36,
"end_time": 195.59,
"duration": 4.23
},
{
"speaker": "speaker_1",
"begin_time": 200.66,
"end_time": 203.28,
"duration": 2.62
},
{
"speaker": "speaker_1",
"begin_time": 203.56,
"end_time": 204.94,
"duration": 1.38
},
{
"speaker": "speaker_2",
"begin_time": 205.22,
"end_time": 206.34,
"duration": 1.12
},
{
"speaker": "speaker_1",
"begin_time": 206.34,
"end_time": 208.59,
"duration": 2.25
},
{
"speaker": "speaker_2",
"begin_time": 208.59,
"end_time": 210.84,
"duration": 2.25
},
{
"speaker": "speaker_1",
"begin_time": 210.84,
"end_time": 213.84,
"duration": 3.0
},
{
"speaker": "speaker_2",
"begin_time": 213.84,
"end_time": 216.09,
"duration": 2.25
},
{
"speaker": "speaker_1",
"begin_time": 216.09,
"end_time": 221.34,
"duration": 5.25
},
{
"speaker": "speaker_2",
"begin_time": 221.34,
"end_time": 225.09,
"duration": 3.75
},
{
"speaker": "speaker_1",
"begin_time": 225.09,
"end_time": 226.59,
"duration": 1.5
},
{
"speaker": "speaker_2",
"begin_time": 226.59,
"end_time": 228.09,
"duration": 1.5
},
{
"speaker": "speaker_1",
"begin_time": 228.09,
"end_time": 231.09,
"duration": 3.0
},
{
"speaker": "speaker_2",
"begin_time": 231.09,
"end_time": 232.59,
"duration": 1.5
},
{
"speaker": "speaker_1",
"begin_time": 232.59,
"end_time": 234.44,
"duration": 1.84
},
{
"speaker": "speaker_1",
"begin_time": 234.99,
"end_time": 236.87,
"duration": 1.88
},
{
"speaker": "speaker_2",
"begin_time": 236.87,
"end_time": 238.37,
"duration": 1.5
},
{
"speaker": "speaker_1",
"begin_time": 238.37,
"end_time": 248.96,
"duration": 10.59
},
{
"speaker": "speaker_1",
"begin_time": 249.24,
"end_time": 252.25,
"duration": 3.01
},
{
"speaker": "speaker_0",
"begin_time": 252.59,
"end_time": 253.4,
"duration": 0.81
},
{
"speaker": "speaker_2",
"begin_time": 253.99,
"end_time": 255.12,
"duration": 1.12
},
{
"speaker": "speaker_0",
"begin_time": 255.12,
"end_time": 255.87,
"duration": 0.75
},
{
"speaker": "speaker_2",
"begin_time": 255.87,
"end_time": 256.62,
"duration": 0.75
},
{
"speaker": "speaker_1",
"begin_time": 256.62,
"end_time": 258.35,
"duration": 1.74
},
{
"speaker": "speaker_0",
"begin_time": 276.76,
"end_time": 277.95,
"duration": 1.19
},
{
"speaker": "speaker_0",
"begin_time": 285.09,
"end_time": 286.96,
"duration": 1.88
},
{
"speaker": "speaker_1",
"begin_time": 286.96,
"end_time": 287.72,
"duration": 0.76
},
{
"speaker": "speaker_2",
"begin_time": 297.92,
"end_time": 299.8,
"duration": 1.88
},
{
"speaker": "speaker_0",
"begin_time": 299.8,
"end_time": 300.55,
"duration": 0.75
},
{
"speaker": "speaker_1",
"begin_time": 300.55,
"end_time": 302.05,
"duration": 1.5
},
{
"speaker": "speaker_0",
"begin_time": 302.05,
"end_time": 305.8,
"duration": 3.75
},
{
"speaker": "speaker_1",
"begin_time": 305.8,
"end_time": 306.55,
"duration": 0.75
},
{
"speaker": "speaker_0",
"begin_time": 306.55,
"end_time": 308.88,
"duration": 2.33
},
{
"speaker": "speaker_0",
"begin_time": 320.97,
"end_time": 323.87,
"duration": 2.9
},
{
"speaker": "speaker_3",
"begin_time": 335.4,
"end_time": 338.77,
"duration": 3.38
},
{
"speaker": "speaker_0",
"begin_time": 338.77,
"end_time": 342.09,
"duration": 3.31
},
{
"speaker": "speaker_0",
"begin_time": 344.76,
"end_time": 345.57,
"duration": 0.81
},
{
"speaker": "speaker_3",
"begin_time": 345.85,
"end_time": 350.66,
"duration": 4.81
},
{
"speaker": "speaker_0",
"begin_time": 352.38,
"end_time": 356.5,
"duration": 4.12
},
{
"speaker": "speaker_1",
"begin_time": 356.5,
"end_time": 357.25,
"duration": 0.75
},
{
"speaker": "speaker_3",
"begin_time": 357.25,
"end_time": 358.0,
"duration": 0.75
},
{
"speaker": "speaker_0",
"begin_time": 358.0,
"end_time": 359.44,
"duration": 1.44
},
{
"speaker": "speaker_2",
"begin_time": 360.43,
"end_time": 362.31,
"duration": 1.88
},
{
"speaker": "speaker_0",
"begin_time": 362.31,
"end_time": 369.81,
"duration": 7.5
},
{
"speaker": "speaker_2",
"begin_time": 369.81,
"end_time": 370.56,
"duration": 0.75
},
{
"speaker": "speaker_3",
"begin_time": 370.56,
"end_time": 372.06,
"duration": 1.5
},
{
"speaker": "speaker_0",
"begin_time": 372.06,
"end_time": 376.66,
"duration": 4.61
},
{
"speaker": "speaker_0",
"begin_time": 376.94,
"end_time": 389.61,
"duration": 12.67
},
{
"speaker": "speaker_0",
"begin_time": 390.19,
"end_time": 398.82,
"duration": 8.63
},
{
"speaker": "speaker_3",
"begin_time": 399.69,
"end_time": 401.67,
"duration": 1.98
},
{
"speaker": "speaker_0",
"begin_time": 401.95,
"end_time": 425.0,
"duration": 23.05
},
{
"speaker": "speaker_0",
"begin_time": 425.32,
"end_time": 430.94,
"duration": 5.62
},
{
"speaker": "speaker_2",
"begin_time": 430.94,
"end_time": 431.69,
"duration": 0.75
},
{
"speaker": "speaker_0",
"begin_time": 431.69,
"end_time": 439.19,
"duration": 7.5
},
{
"speaker": "speaker_3",
"begin_time": 439.19,
"end_time": 440.6,
"duration": 1.41
},
{
"speaker": "speaker_3",
"begin_time": 441.09,
"end_time": 442.21,
"duration": 1.12
},
{
"speaker": "speaker_0",
"begin_time": 442.21,
"end_time": 446.71,
"duration": 4.5
},
{
"speaker": "speaker_3",
"begin_time": 446.71,
"end_time": 447.46,
"duration": 0.75
},
{
"speaker": "speaker_2",
"begin_time": 447.46,
"end_time": 448.21,
"duration": 0.75
},
{
"speaker": "speaker_3",
"begin_time": 448.21,
"end_time": 451.96,
"duration": 3.75
},
{
"speaker": "speaker_0",
"begin_time": 451.96,
"end_time": 452.71,
"duration": 0.75
},
{
"speaker": "speaker_1",
"begin_time": 452.71,
"end_time": 453.46,
"duration": 0.75
},
{
"speaker": "speaker_3",
"begin_time": 453.46,
"end_time": 457.96,
"duration": 4.5
},
{
"speaker": "speaker_0",
"begin_time": 457.96,
"end_time": 475.86,
"duration": 17.9
},
{
"speaker": "speaker_0",
"begin_time": 476.41,
"end_time": 480.54,
"duration": 4.12
},
{
"speaker": "speaker_3",
"begin_time": 480.54,
"end_time": 482.04,
"duration": 1.5
},
{
"speaker": "speaker_2",
"begin_time": 482.04,
"end_time": 488.75,
"duration": 6.71
},
{
"speaker": "speaker_2",
"begin_time": 489.03,
"end_time": 490.15,
"duration": 1.12
},
{
"speaker": "speaker_0",
"begin_time": 490.15,
"end_time": 490.9,
"duration": 0.75
},
{
"speaker": "speaker_3",
"begin_time": 490.9,
"end_time": 492.4,
"duration": 1.5
},
{
"speaker": "speaker_0",
"begin_time": 492.4,
"end_time": 495.4,
"duration": 3.0
},
{
"speaker": "speaker_2",
"begin_time": 495.4,
"end_time": 496.15,
"duration": 0.75
},
{
"speaker": "speaker_0",
"begin_time": 496.15,
"end_time": 496.9,
"duration": 0.75
},
{
"speaker": "speaker_2",
"begin_time": 496.9,
"end_time": 497.65,
"duration": 0.75
},
{
"speaker": "speaker_1",
"begin_time": 497.65,
"end_time": 498.4,
"duration": 0.75
},
{
"speaker": "speaker_3",
"begin_time": 498.4,
"end_time": 499.15,
"duration": 0.75
},
{
"speaker": "speaker_0",
"begin_time": 499.15,
"end_time": 501.4,
"duration": 2.25
},
{
"speaker": "speaker_3",
"begin_time": 501.4,
"end_time": 502.15,
"duration": 0.75
},
{
"speaker": "speaker_0",
"begin_time": 502.15,
"end_time": 514.15,
"duration": 12.0
},
{
"speaker": "speaker_3",
"begin_time": 514.15,
"end_time": 516.4,
"duration": 2.25
},
{
"speaker": "speaker_0",
"begin_time": 516.4,
"end_time": 517.15,
"duration": 0.75
},
{
"speaker": "speaker_3",
"begin_time": 517.15,
"end_time": 520.98,
"duration": 3.83
},
{
"speaker": "speaker_3",
"begin_time": 521.36,
"end_time": 524.15,
"duration": 2.79
},
{
"speaker": "speaker_3",
"begin_time": 525.04,
"end_time": 528.04,
"duration": 3.0
},
{
"speaker": "speaker_3",
"begin_time": 528.69,
"end_time": 529.83,
"duration": 1.14
},
{
"speaker": "speaker_3",
"begin_time": 532.0,
"end_time": 534.62,
"duration": 2.62
},
{
"speaker": "speaker_0",
"begin_time": 534.62,
"end_time": 546.97,
"duration": 12.35
},
{
"speaker": "speaker_0",
"begin_time": 548.95,
"end_time": 551.33,
"duration": 2.38
},
{
"speaker": "speaker_0",
"begin_time": 551.88,
"end_time": 553.0,
"duration": 1.12
},
{
"speaker": "speaker_3",
"begin_time": 553.0,
"end_time": 557.5,
"duration": 4.5
},
{
"speaker": "speaker_0",
"begin_time": 557.5,
"end_time": 563.5,
"duration": 6.0
},
{
"speaker": "speaker_3",
"begin_time": 563.5,
"end_time": 565.0,
"duration": 1.5
},
{
"speaker": "speaker_0",
"begin_time": 565.0,
"end_time": 569.46,
"duration": 4.46
},
{
"speaker": "speaker_3",
"begin_time": 570.57,
"end_time": 571.63,
"duration": 1.06
},
{
"speaker": "speaker_3",
"begin_time": 577.3,
"end_time": 580.67,
"duration": 3.38
},
{
"speaker": "speaker_0",
"begin_time": 580.67,
"end_time": 582.25,
"duration": 1.58
},
{
"speaker": "speaker_0",
"begin_time": 582.59,
"end_time": 586.72,
"duration": 4.12
},
{
"speaker": "speaker_3",
"begin_time": 586.72,
"end_time": 588.97,
"duration": 2.25
},
{
"speaker": "speaker_2",
"begin_time": 588.97,
"end_time": 589.72,
"duration": 0.75
},
{
"speaker": "speaker_0",
"begin_time": 589.72,
"end_time": 596.47,
"duration": 6.75
},
{
"speaker": "speaker_2",
"begin_time": 596.47,
"end_time": 597.22,
"duration": 0.75
},
{
"speaker": "speaker_3",
"begin_time": 597.22,
"end_time": 599.47,
"duration": 2.25
},
{
"speaker": "speaker_0",
"begin_time": 599.47,
"end_time": 599.98,
"duration": 0.51
}
]
}

36
run.bat
View File

@ -1,36 +0,0 @@
@echo off
chcp 65001 >nul
echo ========================================
echo 修复 Windows 路径长度问题
echo ========================================
echo.
REM 设置短路径环境变量
set "MODELSCOPE_CACHE=%~dp0models"
set "FUNASR_MODELS_DIR=%~dp0models"
set "PYTHONLEGACYWINDOWSFSENCODING=1"
REM 创建模型目录
if not exist "models" mkdir models
echo ✅ 环境变量已设置
echo MODELSCOPE_CACHE=%MODELSCOPE_CACHE%
echo FUNASR_MODELS_DIR=%FUNASR_MODELS_DIR%
echo.
REM 检查参数
if "%~1"=="" (
echo 使用方法: fix_path_issue.bat [音频文件路径]
echo 示例: fix_path_issue.bat meeting.wav
pause
exit /b 1
)
echo 🔄 正在运行语音识别...
echo.
REM 使用虚拟环境的 Python 运行
funasr_env\Scripts\python.exe test_asr.py -f "%~1"
echo.
pause

14
server.py
View File

@ -160,7 +160,7 @@ def recognize_single():
min_cluster_size = int(data.get('min_cluster_size', 10)) min_cluster_size = int(data.get('min_cluster_size', 10))
output_format = data.get('format', 'json') output_format = data.get('format', 'json')
filename = secure_filename(file.filename) filename = secure_filename(file.filename) if file.filename is not None else secure_filename("unnamed_file")
task_id = str(uuid.uuid4())[:8] task_id = str(uuid.uuid4())[:8]
audio_path = os.path.join(app.config['UPLOAD_FOLDER'], f"{task_id}_{filename}") audio_path = os.path.join(app.config['UPLOAD_FOLDER'], f"{task_id}_{filename}")
file.save(audio_path) file.save(audio_path)
@ -173,11 +173,7 @@ def recognize_single():
ASR_MODEL_LOADED = True ASR_MODEL_LOADED = True
sentences = service.recognize( sentences = service.recognize(
audio_path, audio_path
use_3d_speaker=use_3d_speaker,
embedding_model=embedding_model,
cluster_threshold=cluster_threshold,
min_cluster_size=min_cluster_size
) )
result = { result = {
@ -194,7 +190,7 @@ def recognize_single():
return jsonify(result) return jsonify(result)
else: else:
srt_path = os.path.join(app.config['RESULT_FOLDER'], f"{task_id}_result.srt") srt_path = os.path.join(app.config['RESULT_FOLDER'], f"{task_id}_result.srt")
service.export_to_srt(sentences, srt_path) service.export_to_srt([s for s in sentences], srt_path)
return send_file(srt_path, as_attachment=True, download_name=f"{task_id}_result.srt") return send_file(srt_path, as_attachment=True, download_name=f"{task_id}_result.srt")
finally: finally:
@ -246,10 +242,6 @@ def recognize_batch():
try: try:
sentences = service.recognize( sentences = service.recognize(
audio_path, audio_path,
use_3d_speaker=use_3d_speaker,
embedding_model=embedding_model,
cluster_threshold=cluster_threshold,
min_cluster_size=min_cluster_size
) )
results.append({ results.append({
'file': os.path.basename(audio_path), 'file': os.path.basename(audio_path),

3
speaker_3D的依赖.md
View File

@ -1,3 +0,0 @@
pip install numpy scipy scikit-learn soundfile kaldiio pyyaml tqdm
pip install umap-learn hdbscan

12
test_3dspeaker.bat
View File

@ -1,12 +0,0 @@
@echo off
echo === Testing 3D-Speaker Import ===
python -c "import sys; print('sys.path:', sys.path[:3])"
echo.
echo === Step 1: Import ===
python -c "import sys; sys.path.insert(0, '..\\3D-Speaker'); from speakerlab.bin.infer_diarization import Diarization3Dspeaker; print('Import OK')"
echo.
echo === Step 2: Init Model ===
python -c "import sys; sys.path.insert(0, '..\\3D-Speaker'); from speakerlab.bin.infer_diarization import Diarization3Dspeaker; m=Diarization3Dspeaker(device='cpu'); print('Init OK')"
echo.
echo === Done ===
pause

153
test_asr.py
View File

@ -1,153 +0,0 @@
"""
FunASR 语音识别测试脚本
支持：句级时间戳、说话人分离（FunASR CAM++ / 3D-Speaker）
"""
import os
import sys
import argparse
from pathlib import Path
def print_banner():
print("=" * 70)
print(" FunASR 语音识别测试工具")
print("=" * 70)
print("功能特性:")
print(" • 句级时间戳（开始时间 - 结束时间）")
print(" • 说话人分离（FunASR CAM++ / 3D-Speaker）")
print(" • 抗噪处理（VAD 语音活动检测）")
print(" • 支持中文、方言、多语言")
print("=" * 70)
print()
def test_single_audio(audio_path: str, model_name: str = "paraformer-zh", use_3d_speaker: bool = False, output_path: str | None = None):
from asr_service import ASRService
if not os.path.exists(audio_path):
print(f"❌ 错误: 文件不存在 - {audio_path}")
return
print(f"🔄 正在初始化模型: {model_name}")
print(f"📝 音频文件: {audio_path}")
if use_3d_speaker:
print(f"🎯 使用 3D-Speaker 替换说话人")
print("-" * 70)
service = ASRService(model_name=model_name)
sentences = service.recognize(audio_path, use_3d_speaker=use_3d_speaker)
print("\n✅ 识别完成！")
print("=" * 70)
print(f"共识别出 {len(sentences)} 句话\n")
for i, sent in enumerate(sentences, 1):
print(f"[{i}] {sent}")
base_name = Path(audio_path).stem
if output_path:
json_path = output_path
srt_path = str(Path(output_path).with_suffix(".srt"))
else:
json_path = f"output/{base_name}_result.json"
srt_path = f"output/{base_name}_result.srt"
service.export_to_json(sentences, json_path)
service.export_to_srt(sentences, srt_path)
print("\n" + "=" * 70)
print("📁 输出文件:")
print(f" • JSON: {json_path}")
print(f" • SRT: {srt_path}")
print("=" * 70)
return sentences
def test_batch(audio_dir: str, model_name: str = "paraformer-zh", use_3d_speaker: bool = False):
from asr_service import ASRService
audio_extensions = {".wav", ".mp3", ".m4a", ".flac", ".ogg", ".wma"}
audio_files = []
for ext in audio_extensions:
audio_files.extend(Path(audio_dir).glob(f"*{ext}"))
if not audio_files:
print(f"❌ 未找到音频文件")
return
print(f"🔄 找到 {len(audio_files)} 个音频文件")
if use_3d_speaker:
print(f"🎯 使用 3D-Speaker 替换说话人")
print("-" * 70)
service = ASRService(model_name=model_name)
for audio_path in audio_files:
print(f"\n处理: {audio_path.name}")
try:
sentences = service.recognize(audio_path, use_3d_speaker=use_3d_speaker)
print(f" ✓ 识别出 {len(sentences)} 句话")
base_name = audio_path.stem
service.export_to_json(sentences, f"output/{base_name}_result.json")
except Exception as e:
print(f" ✗ 失败: {e}")
print("\n✅ 批量处理完成！")
def main():
parser = argparse.ArgumentParser(
description="FunASR 语音识别测试工具",
formatter_class=argparse.RawDescriptionHelpFormatter,
epilog="""
示例用法:
# 识别单个文件（使用内置 CAM++ 说话人分离）
python test_asr.py -f your_audio.wav
# 使用 3D-Speaker 替换说话人（在结果保存前替换）
python test_asr.py -f your_audio.wav --use-3d-speaker
# 指定输出文件
python test_asr.py -f your_audio.wav --use-3d-speaker -o result.json
# 使用 SenseVoice 模型（多语言）
python test_asr.py -f your_audio.wav -m SenseVoice
# 批量识别目录
python test_asr.py -d ./audio_files/
"""
)
parser.add_argument("-f", "--file", help="要识别的音频文件路径")
parser.add_argument("-d", "--directory", help="要批量识别的音频目录")
parser.add_argument("-m", "--model", default="paraformer-zh", choices=["paraformer-zh", "SenseVoice"], help="选择模型")
parser.add_argument("--use-3d-speaker", action="store_true", help="使用 3D-Speaker 替换说话人（在结果保存前替换）")
parser.add_argument("-o", "--output", help="指定输出 JSON 文件路径")
parser.add_argument("--download-sample", action="store_true", help="显示测试音频下载信息")
args = parser.parse_args()
print_banner()
if args.download_sample:
print("📝 请准备测试音频文件")
print("支持的格式: wav, mp3, m4a, flac, ogg, wma")
elif args.file:
test_single_audio(args.file, args.model, args.use_3d_speaker, args.output)
elif args.directory:
test_batch(args.directory, args.model, args.use_3d_speaker)
else:
parser.print_help()
print("\n" + "=" * 70)
print("提示: 使用 -f 指定音频文件，或 -d 指定音频目录")
print(" 使用 --use-3d-speaker 启用 3D-Speaker 替换说话人")
print("=" * 70)
if __name__ == "__main__":
main()

95
test_model_load.py
View File

@ -1,95 +0,0 @@
"""
测试模型加载（不使用多进程）
用于诊断是否是模型本身的问题
"""
import sys
import torch
print("=" * 60)
print("模型加载测试（单进程模式）")
print("=" * 60)
print(f"Python 版本：{sys.version}")
print(f"PyTorch 版本：{torch.__version__}")
print(f"CUDA 可用：{torch.cuda.is_available()}")
if torch.cuda.is_available():
print(f"GPU: {torch.cuda.get_device_name(0)}")
print(f"CUDA 版本：{torch.version.cuda}")
print()
try:
# 测试 ASR 模型
print("=" * 60)
print("测试 1: 加载 ASR 模型 (Paraformer)")
print("=" * 60)
from asr_service import ASRService
asr_service = ASRService(model_name="paraformer-zh", device="auto")
print("✓ ASRService 初始化完成")
asr_service._load_model()
print("✓ ASR 模型加载完成")
print(f" 模型类型：{type(asr_service._model)}")
print()
# 测试说话人分离模型
print("=" * 60)
print("测试 2: 加载说话人分离模型 (3D-Speaker)")
print("=" * 60)
from diarization_service import DiarizationService
diar_service = DiarizationService(
embedding_model="eres2netv2",
device="auto",
cluster_threshold=0.5,
min_cluster_size=10
)
print("✓ DiarizationService 初始化完成")
diar_service._load_model()
print("✓ 说话人分离模型加载完成")
print(f" 模型类型：{type(diar_service.model)}")
print()
# 测试音频处理
print("=" * 60)
print("测试 3: 测试音频处理")
print("=" * 60)
# 检查是否有测试音频
from pathlib import Path
test_audio = Path("test.wav")
if test_audio.exists():
print(f"找到测试音频：{test_audio}")
print("执行 ASR 识别...")
sentences = asr_service.recognize(str(test_audio))
print(f"✓ ASR 识别完成，共 {len(sentences)} 句")
if sentences:
print(f" 第一句：{sentences[0]}")
print("执行说话人分离...")
segments = diar_service.diarize(str(test_audio))
print(f"✓ 说话人分离完成，共 {len(segments)} 个片段")
if segments:
print(f" 第一个片段：{segments[0]}")
else:
print("⚠️ 未找到测试音频 (test.wav)，跳过处理测试")
print()
print("=" * 60)
print("✓ 所有测试通过！模型工作正常")
print("=" * 60)
except Exception as e:
print()
print("=" * 60)
print("✗ 测试失败！")
print("=" * 60)
print(f"错误：{e}")
print()
import traceback
traceback.print_exc()
sys.exit(1)

69
test_staged.py
View File

@ -1,69 +0,0 @@
"""
测试分阶段处理逻辑
"""
from pathlib import Path
import sys
# 导入 main.py 中的函数
from main import (
VIDEO_DIR,
OUTPUT_DIR,
TEMP_DIR,
process_batch_diarization,
process_batch_asr,
get_video_list
)
def test_staged_processing():
"""测试分阶段处理"""
print("=" * 60)
print("分阶段处理测试")
print("=" * 60)
# 获取视频列表（只取前 2 个进行测试）
video_paths = get_video_list(VIDEO_DIR)
if not video_paths:
print("✗ 未找到视频文件")
return
# 只测试前 2 个
test_videos = video_paths[:2]
print(f"测试视频：{len(test_videos)}")
for v in test_videos:
print(f" - {v.name}")
print()
# 阶段 1: 说话人分离
print("=" * 60)
print("阶段 1: 说话人分离")
print("=" * 60)
diar_results = process_batch_diarization(test_videos, max_workers=1)
print(f"\n阶段 1 结果：{len(diar_results)}/{len(test_videos)} 成功")
for video, result_path in diar_results.items():
status = "✓" if result_path else "✗"
print(f" {status} {video.name}: {result_path}")
print()
# 阶段 2: ASR + 合并
print("=" * 60)
print("阶段 2: ASR + 合并")
print("=" * 60)
results = process_batch_asr(test_videos, diar_results, max_workers=1)
print(f"\n阶段 2 结果：{len(results)}/{len(test_videos)} 完成")
for result in results:
status = "✓" if result.get("success") else "✗"
print(f" {status} {Path(result['video']).name}")
if result.get("speaker_counts"):
print(f" 说话人：{result['speaker_counts']}")
print()
print("=" * 60)
print("✓ 测试完成！")
print("=" * 60)
if __name__ == "__main__":
test_staged_processing()