Upload 3 files
Browse files- README.md +192 -15
- app.py +368 -0
- requirements.txt +15 -0
README.md
CHANGED
|
@@ -1,15 +1,192 @@
|
|
| 1 |
-
---
|
| 2 |
-
title:
|
| 3 |
-
emoji:
|
| 4 |
-
colorFrom:
|
| 5 |
-
colorTo:
|
| 6 |
-
sdk: gradio
|
| 7 |
-
sdk_version:
|
| 8 |
-
python_version:
|
| 9 |
-
app_file: app.py
|
| 10 |
-
pinned: false
|
| 11 |
-
|
| 12 |
-
|
| 13 |
-
-
|
| 14 |
-
|
| 15 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
title: SpeechBrain Demo
|
| 3 |
+
emoji: 🧠
|
| 4 |
+
colorFrom: blue
|
| 5 |
+
colorTo: indigo
|
| 6 |
+
sdk: gradio
|
| 7 |
+
sdk_version: "4.44.0"
|
| 8 |
+
python_version: "3.10"
|
| 9 |
+
app_file: app.py
|
| 10 |
+
pinned: false
|
| 11 |
+
---
|
| 12 |
+
|
| 13 |
+
# 🧠 SpeechBrain - 全能语音处理工具包
|
| 14 |
+
|
| 15 |
+
> **SpeechBrain** 是一个基于 PyTorch 的开源语音处理工具包,支持 **语音识别 (ASR)**、**语音合成 (TTS)**、**说话人识别**、**语音增强**、**语音分离**、**语音活动检测 (VAD)** 等多种任务。
|
| 16 |
+
|
| 17 |
+
[](https://www.python.org/)
|
| 18 |
+
[](https://pytorch.org/)
|
| 19 |
+
[](LICENSE)
|
| 20 |
+
[]()
|
| 21 |
+
|
| 22 |
+
---
|
| 23 |
+
|
| 24 |
+
## ✨ 主要功能
|
| 25 |
+
|
| 26 |
+
| 功能 | 描述 | 预训练模型 |
|
| 27 |
+
|------|------|------------|
|
| 28 |
+
| 🎙️ **语音识别 (ASR)** | 将语音转为文字,支持多种语言 | ✅ |
|
| 29 |
+
| 🔊 **语音合成 (TTS)** | 将文字转为自然语音 | ✅ |
|
| 30 |
+
| 👤 **说话人识别** | 识别说话人身份 | ✅ |
|
| 31 |
+
| 🎵 **语音增强** | 降噪、去混响 | ✅ |
|
| 32 |
+
| 🎧 **语音分离** | 分离混合音频中的多个声源 | ✅ |
|
| 33 |
+
| 📊 **语音活动检测** | 检测语音段/非语音段 | ✅ |
|
| 34 |
+
| 📝 **说话人日志** | 谁在什么时候说话 | ✅ |
|
| 35 |
+
| 🌐 **翻译 (ST)** | 语音到文本翻译 | ✅ |
|
| 36 |
+
|
| 37 |
+
---
|
| 38 |
+
|
| 39 |
+
## 🚀 快速开始
|
| 40 |
+
|
| 41 |
+
### 安装
|
| 42 |
+
|
| 43 |
+
```bash
|
| 44 |
+
# 克隆仓库
|
| 45 |
+
git clone https://github.com/zq-1115/demo.git
|
| 46 |
+
cd demo
|
| 47 |
+
|
| 48 |
+
# 安装依赖
|
| 49 |
+
pip install -r requirements.txt
|
| 50 |
+
```
|
| 51 |
+
|
| 52 |
+
### 30秒体验
|
| 53 |
+
|
| 54 |
+
```python
|
| 55 |
+
from speechbrain.inference import EncoderDecoderASR
|
| 56 |
+
|
| 57 |
+
# 下载预训练模型并转录音频
|
| 58 |
+
asr_model = EncoderDecoderASR.from_hparams(
|
| 59 |
+
source="speechbrain/asr-crdnn-rnnlm-librispeech",
|
| 60 |
+
savedir="pretrained_models/asr-crdnn-rnnlm-librispeech",
|
| 61 |
+
)
|
| 62 |
+
|
| 63 |
+
# 转录音频文件
|
| 64 |
+
text = asr_model.transcribe_file("demo_audio.wav")
|
| 65 |
+
print(f"识别结果: {text}")
|
| 66 |
+
```
|
| 67 |
+
|
| 68 |
+
---
|
| 69 |
+
|
| 70 |
+
## 📂 项目结构
|
| 71 |
+
|
| 72 |
+
```
|
| 73 |
+
speechbrain-develop/
|
| 74 |
+
├── speechbrain/ # 核心库代码
|
| 75 |
+
│ ├── inference/ # 推理接口 (ASR/TTS/VAD等)
|
| 76 |
+
│ ├── dataio/ # 数据加载与处理
|
| 77 |
+
│ ├── decoders/ # 解码器 (CTC/Seq2Seq等)
|
| 78 |
+
│ ├── augment/ # 数据增强
|
| 79 |
+
│ ├── alignment/ # 对齐工具
|
| 80 |
+
│ ├── integrations/ # 集成模块 (HuggingFace/K2等)
|
| 81 |
+
│ └── utils/ # 工具函数
|
| 82 |
+
├── tests/ # 测试与示例音频
|
| 83 |
+
├── demo.py # 🎯 交互式演示脚本
|
| 84 |
+
├── requirements.txt # 依赖列表
|
| 85 |
+
└── setup.py # 安装配置
|
| 86 |
+
```
|
| 87 |
+
|
| 88 |
+
---
|
| 89 |
+
|
| 90 |
+
## 🎮 运行演示
|
| 91 |
+
|
| 92 |
+
```bash
|
| 93 |
+
python demo.py
|
| 94 |
+
```
|
| 95 |
+
|
| 96 |
+
演示程序提供交互式菜单,可以选择:
|
| 97 |
+
1. **语音识别 (ASR)** — 上传音频,输出文字
|
| 98 |
+
2. **说话人识别** — 识别说话人
|
| 99 |
+
3. **语音活动检测 (VAD)** — 检测语音片段
|
| 100 |
+
|
| 101 |
+
---
|
| 102 |
+
|
| 103 |
+
## 🔧 使用示例
|
| 104 |
+
|
| 105 |
+
### 1. 语音识别 (ASR)
|
| 106 |
+
|
| 107 |
+
```python
|
| 108 |
+
from speechbrain.inference import EncoderDecoderASR
|
| 109 |
+
|
| 110 |
+
asr = EncoderDecoderASR.from_hparams(
|
| 111 |
+
source="speechbrain/asr-crdnn-rnnlm-librispeech",
|
| 112 |
+
savedir="pretrained_models/asr",
|
| 113 |
+
)
|
| 114 |
+
result = asr.transcribe_file("audio.wav")
|
| 115 |
+
print(result) # "MY FATHER HAS REVEALED THE CULPRIT'S NAME"
|
| 116 |
+
```
|
| 117 |
+
|
| 118 |
+
### 2. 说话人识别
|
| 119 |
+
|
| 120 |
+
```python
|
| 121 |
+
from speechbrain.inference import SpeakerRecognition
|
| 122 |
+
|
| 123 |
+
spk_rec = SpeakerRecognition.from_hparams(
|
| 124 |
+
source="speechbrain/spkrec-ecapa-voxceleb",
|
| 125 |
+
savedir="pretrained_models/spkrec",
|
| 126 |
+
)
|
| 127 |
+
score, prediction = spk_rec.verify_files("spk1.wav", "spk2.wav")
|
| 128 |
+
print(f"是同一个人吗? {'是' if prediction else '否'}")
|
| 129 |
+
```
|
| 130 |
+
|
| 131 |
+
### 3. 语音活动检测 (VAD)
|
| 132 |
+
|
| 133 |
+
```python
|
| 134 |
+
from speechbrain.inference import VAD
|
| 135 |
+
|
| 136 |
+
vad = VAD.from_hparams(
|
| 137 |
+
source="speechbrain/vad-crdnn-libriparty",
|
| 138 |
+
savedir="pretrained_models/vad",
|
| 139 |
+
)
|
| 140 |
+
boundaries = vad.get_speech_segments("audio.wav")
|
| 141 |
+
for start, end in boundaries:
|
| 142 |
+
print(f"语音段: {start:.2f}s - {end:.2f}s")
|
| 143 |
+
```
|
| 144 |
+
|
| 145 |
+
### 4. 语音合成 (TTS)
|
| 146 |
+
|
| 147 |
+
```python
|
| 148 |
+
from speechbrain.inference import Tacotron2
|
| 149 |
+
from speechbrain.inference import HIFIGAN
|
| 150 |
+
|
| 151 |
+
tacotron2 = Tacotron2.from_hparams(
|
| 152 |
+
source="speechbrain/tts-tacotron2-ljspeech",
|
| 153 |
+
savedir="pretrained_models/tts-tacotron2",
|
| 154 |
+
)
|
| 155 |
+
hifi_gan = HIFIGAN.from_hparams(
|
| 156 |
+
source="speechbrain/tts-hifigan-ljspeech",
|
| 157 |
+
savedir="pretrained_models/tts-hifigan",
|
| 158 |
+
)
|
| 159 |
+
|
| 160 |
+
# 文字转语音
|
| 161 |
+
mel_output, mel_length, alignment = tacotron2.encode_text("Hello world")
|
| 162 |
+
waveforms = hifi_gan.decode_batch(mel_output)
|
| 163 |
+
torchaudio.save("output.wav", waveforms.squeeze(1), 22050)
|
| 164 |
+
```
|
| 165 |
+
|
| 166 |
+
---
|
| 167 |
+
|
| 168 |
+
## 📦 依赖
|
| 169 |
+
|
| 170 |
+
- **Python** >= 3.8
|
| 171 |
+
- **PyTorch** >= 2.1.0
|
| 172 |
+
- **torchaudio** >= 2.1.0
|
| 173 |
+
- **transformers** >= 4.30.0
|
| 174 |
+
- 完整列表见 [requirements.txt](requirements.txt)
|
| 175 |
+
|
| 176 |
+
---
|
| 177 |
+
|
| 178 |
+
## 🤝 贡献
|
| 179 |
+
|
| 180 |
+
欢迎提交 Issue 和 Pull Request!
|
| 181 |
+
|
| 182 |
+
---
|
| 183 |
+
|
| 184 |
+
## 📄 许可证
|
| 185 |
+
|
| 186 |
+
本项目基于 Apache 2.0 许可证开源。详见 [LICENSE](LICENSE)。
|
| 187 |
+
|
| 188 |
+
---
|
| 189 |
+
|
| 190 |
+
## ⭐ Star History
|
| 191 |
+
|
| 192 |
+
如果这个项目对你有帮助,请给一个 ⭐ Star 支持一下!
|
app.py
ADDED
|
@@ -0,0 +1,368 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
SpeechBrain Hugging Face Spaces Demo
|
| 3 |
+
=====================================
|
| 4 |
+
基于 Gradio 的交互式演示,用于 Hugging Face Spaces 部署。
|
| 5 |
+
|
| 6 |
+
功能:
|
| 7 |
+
1. 语音识别 (ASR) — 将语音转录为文字
|
| 8 |
+
2. 说话人验证 — 判断两段音频是否为同一人
|
| 9 |
+
3. 语音活动检测 (VAD) — 检测音频中的语音片段
|
| 10 |
+
4. 说话人嵌入提取 — 提取声纹特征向量
|
| 11 |
+
"""
|
| 12 |
+
|
| 13 |
+
import os
|
| 14 |
+
import sys
|
| 15 |
+
import warnings
|
| 16 |
+
|
| 17 |
+
warnings.filterwarnings("ignore")
|
| 18 |
+
|
| 19 |
+
# ============================================================
|
| 20 |
+
# Gradio / HuggingFace Spaces 兼容性修复
|
| 21 |
+
# ============================================================
|
| 22 |
+
|
| 23 |
+
# --- 修复 1: gradio_client 的 boolean schema 问题 ---
|
| 24 |
+
# gradio_client 在处理 additionalProperties: true 时,
|
| 25 |
+
# 将 boolean True 作为 schema 传入,导致 "const" in schema 崩溃
|
| 26 |
+
try:
|
| 27 |
+
import gradio_client.utils as _gc_utils
|
| 28 |
+
|
| 29 |
+
_original_get_type = _gc_utils.get_type
|
| 30 |
+
|
| 31 |
+
def _patched_get_type(schema):
|
| 32 |
+
if isinstance(schema, bool):
|
| 33 |
+
return "boolean"
|
| 34 |
+
return _original_get_type(schema)
|
| 35 |
+
|
| 36 |
+
_gc_utils.get_type = _patched_get_type
|
| 37 |
+
|
| 38 |
+
_original_json_schema = _gc_utils._json_schema_to_python_type
|
| 39 |
+
|
| 40 |
+
def _patched_json_schema(schema, defs):
|
| 41 |
+
if isinstance(schema, bool):
|
| 42 |
+
return "boolean"
|
| 43 |
+
return _original_json_schema(schema, defs)
|
| 44 |
+
|
| 45 |
+
_gc_utils._json_schema_to_python_type = _patched_json_schema
|
| 46 |
+
except ImportError:
|
| 47 |
+
pass
|
| 48 |
+
|
| 49 |
+
# --- 修复 2: Jinja2 缓存 key 不可哈希问题 ---
|
| 50 |
+
# 某些 gradio 版本传给 Jinja2 的 globals 包含不可哈希 dict
|
| 51 |
+
try:
|
| 52 |
+
import jinja2.environment as _jinja_env
|
| 53 |
+
|
| 54 |
+
_original_load_template = _jinja_env.Environment._load_template
|
| 55 |
+
|
| 56 |
+
def _patched_load_template(self, name, globals):
|
| 57 |
+
try:
|
| 58 |
+
return _original_load_template(self, name, globals)
|
| 59 |
+
except TypeError:
|
| 60 |
+
# 缓存 key 不可哈希时,传入空 dict(而非 None),
|
| 61 |
+
# 因为 Jinja2 会将 None 替换为 self.globals,而后者也可能包含不可哈希类型
|
| 62 |
+
return _original_load_template(self, name, {})
|
| 63 |
+
|
| 64 |
+
_jinja_env.Environment._load_template = _patched_load_template
|
| 65 |
+
except ImportError:
|
| 66 |
+
pass
|
| 67 |
+
|
| 68 |
+
# ============================================================
|
| 69 |
+
# Windows 兼容性修复
|
| 70 |
+
# ============================================================
|
| 71 |
+
if sys.platform == "win32":
|
| 72 |
+
import types
|
| 73 |
+
for _dep in ["k2", "flair", "numba", "spacy"]:
|
| 74 |
+
if _dep not in sys.modules:
|
| 75 |
+
m = types.ModuleType(_dep)
|
| 76 |
+
m.__path__ = []
|
| 77 |
+
sys.modules[_dep] = m
|
| 78 |
+
|
| 79 |
+
import torch
|
| 80 |
+
import gradio as gr
|
| 81 |
+
|
| 82 |
+
# ============================================================
|
| 83 |
+
# 模型缓存目录(HuggingFace Spaces 会自动缓存到 /data 或默认目录)
|
| 84 |
+
# ============================================================
|
| 85 |
+
MODEL_DIR = os.path.join(os.path.dirname(__file__), "pretrained_models")
|
| 86 |
+
|
| 87 |
+
_asr_model = None
|
| 88 |
+
_spk_model = None
|
| 89 |
+
_vad_model = None
|
| 90 |
+
|
| 91 |
+
|
| 92 |
+
def get_device():
|
| 93 |
+
"""获取可用设备"""
|
| 94 |
+
return "cuda" if torch.cuda.is_available() else "cpu"
|
| 95 |
+
|
| 96 |
+
|
| 97 |
+
# ============================================================
|
| 98 |
+
# 模型加载(懒加载,首次使用时才下载)
|
| 99 |
+
# ============================================================
|
| 100 |
+
|
| 101 |
+
def get_asr_model():
|
| 102 |
+
"""语音识别模型 (CRDNN + RNNLM)"""
|
| 103 |
+
global _asr_model
|
| 104 |
+
if _asr_model is None:
|
| 105 |
+
from speechbrain.inference.ASR import EncoderDecoderASR
|
| 106 |
+
print("[INFO] 正在加载 ASR 模型...")
|
| 107 |
+
_asr_model = EncoderDecoderASR.from_hparams(
|
| 108 |
+
source="speechbrain/asr-crdnn-rnnlm-librispeech",
|
| 109 |
+
savedir=os.path.join(MODEL_DIR, "asr-crdnn-rnnlm-librispeech"),
|
| 110 |
+
run_opts={"device": get_device()},
|
| 111 |
+
)
|
| 112 |
+
return _asr_model
|
| 113 |
+
|
| 114 |
+
|
| 115 |
+
def get_speaker_model():
|
| 116 |
+
"""说话人识别模型 (ECAPA-TDNN)"""
|
| 117 |
+
global _spk_model
|
| 118 |
+
if _spk_model is None:
|
| 119 |
+
from speechbrain.inference.speaker import SpeakerRecognition
|
| 120 |
+
print("[INFO] 正在加载说话人识别模型...")
|
| 121 |
+
_spk_model = SpeakerRecognition.from_hparams(
|
| 122 |
+
source="speechbrain/spkrec-ecapa-voxceleb",
|
| 123 |
+
savedir=os.path.join(MODEL_DIR, "spkrec-ecapa-voxceleb"),
|
| 124 |
+
run_opts={"device": get_device()},
|
| 125 |
+
)
|
| 126 |
+
return _spk_model
|
| 127 |
+
|
| 128 |
+
|
| 129 |
+
def get_vad_model():
|
| 130 |
+
"""语音活动检测模型 (CRDNN)"""
|
| 131 |
+
global _vad_model
|
| 132 |
+
if _vad_model is None:
|
| 133 |
+
from speechbrain.inference.VAD import VAD
|
| 134 |
+
print("[INFO] 正在加载 VAD 模型...")
|
| 135 |
+
_vad_model = VAD.from_hparams(
|
| 136 |
+
source="speechbrain/vad-crdnn-libriparty",
|
| 137 |
+
savedir=os.path.join(MODEL_DIR, "vad-crdnn-libriparty"),
|
| 138 |
+
run_opts={"device": get_device()},
|
| 139 |
+
)
|
| 140 |
+
return _vad_model
|
| 141 |
+
|
| 142 |
+
|
| 143 |
+
# ============================================================
|
| 144 |
+
# 功能函数
|
| 145 |
+
# ============================================================
|
| 146 |
+
|
| 147 |
+
def transcribe_audio(audio_file):
|
| 148 |
+
"""语音识别:将音频转为文字"""
|
| 149 |
+
if audio_file is None:
|
| 150 |
+
return "⚠️ 请先上传音频文件"
|
| 151 |
+
|
| 152 |
+
try:
|
| 153 |
+
asr = get_asr_model()
|
| 154 |
+
text = asr.transcribe_file(audio_file)
|
| 155 |
+
return f"📝 **识别结果:**\n\n> {text}"
|
| 156 |
+
except Exception as e:
|
| 157 |
+
return f"❌ 识别失败:{str(e)}"
|
| 158 |
+
|
| 159 |
+
|
| 160 |
+
def verify_speakers(audio1, audio2):
|
| 161 |
+
"""说话人验证:判断两段音频是否为同一说话人"""
|
| 162 |
+
if audio1 is None or audio2 is None:
|
| 163 |
+
return "⚠️ 请上传两段音频文件"
|
| 164 |
+
|
| 165 |
+
try:
|
| 166 |
+
spk = get_speaker_model()
|
| 167 |
+
score, prediction = spk.verify_files(audio1, audio2)
|
| 168 |
+
|
| 169 |
+
similarity = score.item()
|
| 170 |
+
is_same = "✅ 是同一说话人" if prediction else "❌ 不是同一说话人"
|
| 171 |
+
|
| 172 |
+
if similarity > 0.5:
|
| 173 |
+
level = "🟢 高"
|
| 174 |
+
elif similarity > 0.0:
|
| 175 |
+
level = "🟡 中"
|
| 176 |
+
else:
|
| 177 |
+
level = "🔴 低"
|
| 178 |
+
|
| 179 |
+
return (
|
| 180 |
+
f"## {is_same}\n\n"
|
| 181 |
+
f"| 指标 | 值 |\n"
|
| 182 |
+
f"|------|----|\n"
|
| 183 |
+
f"| **相似度得分** | {similarity:.4f} |\n"
|
| 184 |
+
f"| **置信程度** | {level} |\n"
|
| 185 |
+
f"| **判定阈值** | 0.0(高于此值判定为同一人) |"
|
| 186 |
+
)
|
| 187 |
+
except Exception as e:
|
| 188 |
+
return f"❌ 验证失败:{str(e)}"
|
| 189 |
+
|
| 190 |
+
|
| 191 |
+
def detect_speech(audio_file):
|
| 192 |
+
"""语音活动检测:标记音频中的语音段"""
|
| 193 |
+
if audio_file is None:
|
| 194 |
+
return "⚠️ 请先上传音频文件"
|
| 195 |
+
|
| 196 |
+
try:
|
| 197 |
+
vad = get_vad_model()
|
| 198 |
+
boundaries = vad.get_speech_segments(audio_file)
|
| 199 |
+
|
| 200 |
+
if len(boundaries) == 0:
|
| 201 |
+
return "🔇 未检测到语音"
|
| 202 |
+
|
| 203 |
+
lines = [
|
| 204 |
+
"## 🎯 检测结果\n",
|
| 205 |
+
f"| 片段 | 开始 | 结束 | 时长 |",
|
| 206 |
+
f"|------|------|------|------|",
|
| 207 |
+
]
|
| 208 |
+
|
| 209 |
+
total_duration = 0
|
| 210 |
+
for i, (start_t, end_t) in enumerate(boundaries, 1):
|
| 211 |
+
duration = end_t - start_t
|
| 212 |
+
total_duration += duration
|
| 213 |
+
lines.append(
|
| 214 |
+
f"| {i} | {start_t:.2f}s | {end_t:.2f}s | {duration:.2f}s |"
|
| 215 |
+
)
|
| 216 |
+
|
| 217 |
+
lines.append(
|
| 218 |
+
f"\n📊 **共 {len(boundaries)} 个语音片段,总时长 {total_duration:.2f}s**"
|
| 219 |
+
)
|
| 220 |
+
return "\n".join(lines)
|
| 221 |
+
except Exception as e:
|
| 222 |
+
return f"❌ 检测失败:{str(e)}"
|
| 223 |
+
|
| 224 |
+
|
| 225 |
+
def extract_speaker_embedding(audio_file):
|
| 226 |
+
"""提取说话人嵌入向量(声纹特征)"""
|
| 227 |
+
if audio_file is None:
|
| 228 |
+
return "⚠️ 请先上传音频文件"
|
| 229 |
+
|
| 230 |
+
try:
|
| 231 |
+
spk = get_speaker_model()
|
| 232 |
+
embedding = spk.encode_file(audio_file)
|
| 233 |
+
shape = tuple(embedding.shape)
|
| 234 |
+
return (
|
| 235 |
+
f"✅ **提取成功!**\n\n"
|
| 236 |
+
f"- 嵌入向量维度:`{shape}`\n"
|
| 237 |
+
f"- 前 10 个值:`{embedding[0, :10].tolist()}`\n"
|
| 238 |
+
f"- 用途:说话人聚类、验证、识别等"
|
| 239 |
+
)
|
| 240 |
+
except Exception as e:
|
| 241 |
+
return f"❌ 提取失败:{str(e)}"
|
| 242 |
+
|
| 243 |
+
|
| 244 |
+
# ============================================================
|
| 245 |
+
# Gradio 界面
|
| 246 |
+
# ============================================================
|
| 247 |
+
|
| 248 |
+
def create_demo():
|
| 249 |
+
device = get_device()
|
| 250 |
+
device_name = "GPU 🚀" if device == "cuda" else "CPU ⚡"
|
| 251 |
+
|
| 252 |
+
with gr.Blocks(
|
| 253 |
+
title="SpeechBrain Demo — 语音处理工具箱",
|
| 254 |
+
theme=gr.themes.Soft(),
|
| 255 |
+
css="""
|
| 256 |
+
.output-markdown { font-size: 16px; }
|
| 257 |
+
footer { visibility: hidden; }
|
| 258 |
+
""",
|
| 259 |
+
) as demo:
|
| 260 |
+
gr.Markdown(
|
| 261 |
+
f"""
|
| 262 |
+
# 🧠 SpeechBrain 语音处理 Demo
|
| 263 |
+
### 基于 PyTorch 的全能语音处理工具包 | 运行设备:**{device_name}**
|
| 264 |
+
|
| 265 |
+
上传音频文件,体验语音识别、说话人验证、语音活动检测等功能。
|
| 266 |
+
首次使用会自动从 HuggingFace 下载预训练模型。
|
| 267 |
+
"""
|
| 268 |
+
)
|
| 269 |
+
|
| 270 |
+
with gr.Tabs():
|
| 271 |
+
# ── Tab 1: 语音识别 ──
|
| 272 |
+
with gr.Tab("🎙️ 语音识别 (ASR)"):
|
| 273 |
+
gr.Markdown("上传英文音频,自动转录为文字。支持上传文件或麦克风录音。")
|
| 274 |
+
with gr.Row():
|
| 275 |
+
with gr.Column(scale=1):
|
| 276 |
+
audio_asr = gr.Audio(
|
| 277 |
+
label="上传音频",
|
| 278 |
+
type="filepath",
|
| 279 |
+
sources=["upload", "microphone"],
|
| 280 |
+
)
|
| 281 |
+
btn_asr = gr.Button("🔍 开始识别", variant="primary", size="lg")
|
| 282 |
+
with gr.Column(scale=1):
|
| 283 |
+
output_asr = gr.Markdown(value="等待上传音频...")
|
| 284 |
+
|
| 285 |
+
btn_asr.click(
|
| 286 |
+
fn=transcribe_audio,
|
| 287 |
+
inputs=audio_asr,
|
| 288 |
+
outputs=output_asr,
|
| 289 |
+
)
|
| 290 |
+
|
| 291 |
+
# ── Tab 2: 说话人验证 ──
|
| 292 |
+
with gr.Tab("👤 说话人验证"):
|
| 293 |
+
gr.Markdown("上传两段音频,判断是否为同一个人说话。")
|
| 294 |
+
with gr.Row():
|
| 295 |
+
with gr.Column():
|
| 296 |
+
audio_spk1 = gr.Audio(label="音频 1", type="filepath")
|
| 297 |
+
with gr.Column():
|
| 298 |
+
audio_spk2 = gr.Audio(label="音频 2", type="filepath")
|
| 299 |
+
btn_spk = gr.Button("🔍 开始验证", variant="primary", size="lg")
|
| 300 |
+
output_spk = gr.Markdown(value="等待上传两段音频...")
|
| 301 |
+
|
| 302 |
+
btn_spk.click(
|
| 303 |
+
fn=verify_speakers,
|
| 304 |
+
inputs=[audio_spk1, audio_spk2],
|
| 305 |
+
outputs=output_spk,
|
| 306 |
+
)
|
| 307 |
+
|
| 308 |
+
# ── Tab 3: 语音活动检测 ──
|
| 309 |
+
with gr.Tab("📊 语音检测 (VAD)"):
|
| 310 |
+
gr.Markdown("检测音频中哪些时间段有语音活动。")
|
| 311 |
+
with gr.Row():
|
| 312 |
+
with gr.Column(scale=1):
|
| 313 |
+
audio_vad = gr.Audio(
|
| 314 |
+
label="上传音频",
|
| 315 |
+
type="filepath",
|
| 316 |
+
sources=["upload", "microphone"],
|
| 317 |
+
)
|
| 318 |
+
btn_vad = gr.Button("🔍 开始检测", variant="primary", size="lg")
|
| 319 |
+
with gr.Column(scale=1):
|
| 320 |
+
output_vad = gr.Markdown(value="等待上传音频...")
|
| 321 |
+
|
| 322 |
+
btn_vad.click(
|
| 323 |
+
fn=detect_speech,
|
| 324 |
+
inputs=audio_vad,
|
| 325 |
+
outputs=output_vad,
|
| 326 |
+
)
|
| 327 |
+
|
| 328 |
+
# ── Tab 4: 说话人嵌入 ──
|
| 329 |
+
with gr.Tab("🧬 声纹提取"):
|
| 330 |
+
gr.Markdown("提取说话人的声纹特征向量(Embedding),可用于下游任务。")
|
| 331 |
+
with gr.Row():
|
| 332 |
+
with gr.Column(scale=1):
|
| 333 |
+
audio_emb = gr.Audio(label="上传音频", type="filepath")
|
| 334 |
+
btn_emb = gr.Button("🔍 提取嵌入", variant="primary", size="lg")
|
| 335 |
+
with gr.Column(scale=1):
|
| 336 |
+
output_emb = gr.Markdown(value="等待上传音频...")
|
| 337 |
+
|
| 338 |
+
btn_emb.click(
|
| 339 |
+
fn=extract_speaker_embedding,
|
| 340 |
+
inputs=audio_emb,
|
| 341 |
+
outputs=output_emb,
|
| 342 |
+
)
|
| 343 |
+
|
| 344 |
+
gr.Markdown(
|
| 345 |
+
"""
|
| 346 |
+
---
|
| 347 |
+
### 📚 相关链接
|
| 348 |
+
- [SpeechBrain GitHub](https://github.com/speechbrain/speechbrain)
|
| 349 |
+
- [HuggingFace 模型库](https://huggingface.co/speechbrain)
|
| 350 |
+
- 模型:ASR (CRDNN+RNNLM) | 说话人识别 (ECAPA-TDNN) | VAD (CRDNN)
|
| 351 |
+
"""
|
| 352 |
+
)
|
| 353 |
+
|
| 354 |
+
return demo
|
| 355 |
+
|
| 356 |
+
|
| 357 |
+
# ============================================================
|
| 358 |
+
# HuggingFace Spaces 入口
|
| 359 |
+
# ============================================================
|
| 360 |
+
if __name__ == "__main__":
|
| 361 |
+
demo = create_demo()
|
| 362 |
+
# HuggingFace Spaces 会自动注入环境变量,本地开发时绑定 0.0.0.0:7860
|
| 363 |
+
server_name = os.environ.get("GRADIO_SERVER_NAME", "0.0.0.0")
|
| 364 |
+
server_port = int(os.environ.get("GRADIO_SERVER_PORT", 7860))
|
| 365 |
+
demo.launch(
|
| 366 |
+
server_name=server_name,
|
| 367 |
+
server_port=server_port,
|
| 368 |
+
)
|
requirements.txt
ADDED
|
@@ -0,0 +1,15 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
speechbrain
|
| 2 |
+
gradio>=4.44.0,<5.0.0
|
| 3 |
+
huggingface_hub>=0.8.0,<1.0
|
| 4 |
+
hyperpyyaml>=0.0.1
|
| 5 |
+
joblib>=0.14.1
|
| 6 |
+
numpy>=1.17.0
|
| 7 |
+
packaging
|
| 8 |
+
requests>=2.20.0
|
| 9 |
+
scipy>=1.4.1
|
| 10 |
+
sentencepiece>=0.1.91
|
| 11 |
+
soundfile>=0.12.1
|
| 12 |
+
torch>=2.1.0
|
| 13 |
+
torchaudio>=2.1.0
|
| 14 |
+
tqdm>=4.42.0
|
| 15 |
+
transformers>=4.30.0
|