Heehobino/nihao-zhenzhen-kws
By Heehobino · ↓ 3 · ♥ 0 · Updated 2026-02-04
What's in this repo
This page lists the files in Heehobino/nihao-zhenzhen-kws, including decoder.onnx, encoder.onnx, joiner.onnx and more.
- 6 ONNX models (.onnx files)
- 17 files, including 6 .onnx, 5 .txt, 2 .py, 1 .md, 1 .json
- Last updated 2026-02-04
ONNX files
| File | Size | Download |
|---|---|---|
| model/decoder.onnx | — | Download |
| model/encoder.onnx | — | Download |
| model/joiner.onnx | — | Download |
| model_int8/decoder.onnx | — | Download |
| model_int8/encoder.onnx | — | Download |
| model_int8/joiner.onnx | — | Download |
How to use
These ONNX files can be loaded with ONNX Runtime. To inspect a model graph, upload the file to the Netron viewer. Download links point to the original files on Hugging Face.
README
你好真真 - 中文关键词检测模型
中文唤醒词/关键词检测模型,用于检测 "你好真真" (Ni Hao Zhen Zhen) 关键词。
模型变体
提供FP32和INT8两个版本:
| 版本 | 大小 | 阈值 | F1 | FAR | FRR | RTF |
|---|---|---|---|---|---|---|
| FP32 (默认) | 13MB | 0.52 | 89.47% | 5.64% | 2.30% | 0.0185 |
| INT8 | 5MB | 0.46 | 89.01% | 5.96% | 2.30% | 0.0140 |
- FP32: 精度更高,推荐用于服务器端
- INT8: 体积小62%,速度快24%,推荐用于嵌入式/移动端
安装
pip install -r requirements.txt
或手动安装依赖:
pip install sherpa-onnx sounddevice soundfile numpy
快速开始
实时麦克风检测
cd examples
python realtime_detection.py
运行后对着麦克风说 "你好真真" 即可看到检测结果。
Python API
from inference import load_model
# 加载FP32模型(默认)
detector = load_model()
# 或加载INT8模型(更小更快)
detector = load_model(variant="int8")
# 检测音频文件
result = detector.detect("path/to/audio.wav")
print(f"检测到关键词: {result['detected']}")
print(f"关键词: {result['keyword']}")
流式检测
import numpy as np
from inference import load_model
detector = load_model()
stream = detector.create_stream()
# 模拟流式音频输入
chunk_size = 1600 # 100ms @ 16kHz
audio_data = np.random.randn(chunk_size).astype(np.float32)
stream.accept_waveform(16000, audio_data.tolist())
while detector._kws.is_ready(stream):
detector._kws.decode_stream(stream)
result = detector._kws.get_result(stream)
print(result)
目录结构
nihao-zhenzhen-kws/
├── model/ # FP32模型 (13MB)
│ ├── encoder.onnx
│ ├── decoder.onnx
│ ├── joiner.onnx
│ ├── tokens.txt
│ └── keywords.txt
├── model_int8/ # INT8模型 (5MB)
│ ├── encoder.onnx
│ ├── decoder.onnx
│ ├── joiner.onnx
│ ├── tokens.txt
│ └── keywords.txt
├── examples/
│ └── realtime_detection.py
├── config.json
├── inference.py
├── requirements.txt
└── README.md
性能指标
在406个真实人声测试样本上的评估结果(63个正样本 "你好真真",28个相似词 "你好珍珍/娟娟",315个负样本):
FP32模型
- F1分数: 89.47%
- 准确率: 95.07%
- 误报率 (FAR): 5.64%
- 漏检率 (FRR): 2.30%
- RTF: 0.0185
INT8模型
- F1分数: 89.01%
- 准确率: 94.33%
- 误报率 (FAR): 5.96%
- 漏检率 (FRR): 2.30%
- RTF: 0.0140
License
MIT License
致谢
- sherpa-onnx - 语音识别推理框架
- icefall - 模型训练框架
Model info
audio-classification
sherpa-onnx
sherpa-onnx
onnx
keyword-spotting
wake-word
speech-recognition
zipformer
streaming
int8
quantization
audio-classification
zh
license:mit
↓ 3
♥ 0
Created: 2026-02-03
Updated: 2026-02-04
View on Hugging FaceFiles (17)
.gitattributes
LICENSE
README.md
config.json
examples/realtime_detection.py
inference.py
model/decoder.onnx
ONNX
model/encoder.onnx
ONNX
model/joiner.onnx
ONNX
model/keywords.txt
model/tokens.txt
model_int8/decoder.onnx
ONNX
model_int8/encoder.onnx
ONNX
model_int8/joiner.onnx
ONNX
model_int8/keywords.txt
model_int8/tokens.txt
requirements.txt