[2025/11] WASM build guide and browser test demo are now available in lib/Web and examples.
[2025/11] We supported Python inference with ONNX model on Linux, macOS thanks to Guy Nicholson!
[2025/11] We supported Golang on Linux, macOS and Windows with usage of the prebuilt-libs thanks to hylarucoder!
[2025/11] We supported Java on Linux, macOS, Windows, Android with usage of the prebuilt-libs thanks to ZhangYang!
[2025/07] 🎉 Exciting news! TEN VAD is now integrated into k2-fsa/sherpa-onnx, thanks to the fantastic work by Fangjun Kuang! You can now achieve more precise speech segment extraction and enjoy an enhanced ASR experience! Refer to the documentation and give it a try!
[2025/07] We support Python inference on macOS and Windows with usage of the prebuilt-lib!
[2025/06] We finally released and open-sourced the ONNX model and the corresponding preprocessing code! Now you can deploy TEN VAD on any platform and any hardware architecture!
[2025/06] We are excited to announce the release of WASM+JS for Web WASM Support.
TEN is a collection of open-source projects for building real-time, multimodal conversational voice agents. It includes TEN Framework , TEN VAD , TEN Turn Detection , TEN Agent, TMAN Designer, and TEN Portal , all fully open-source.
TEN VAD is a real-time voice activity detection system designed for enterprise use, providing accurate frame-level speech activity detection. It shows superior precision compared to both WebRTC VAD and Silero VAD, which are commonly used in the industry. Additionally, TEN VAD offers lower computational complexity and reduced memory usage compared to Silero VAD. Meanwhile, the architecture's temporal efficiency enables rapid voice activity detection, significantly reducing end-to-end response and turn detection latency in conversational AI systems.
Key Features
1. High-Performance:
The precision-recall curves comparing the performance of WebRTC VAD (pitch-based), Silero VAD, and TEN VAD are shown below. The evaluation is conducted on the precisely manually annotated testset. The audio files are from librispeech, gigaspeech, DNS Challenge etc. As demonstrated, TEN VAD achieves the best performance. Additionally, cross-validation experiments conducted on large internal real-world datasets demonstrate the reproducibility of these findings. The testset with annotated labels is released in directory "testset" of this repository.
Note that the default threshold of 0.5 is used to generate binary speech indicators (0 for non-speech signal, 1 for speech signal). This threshold needs to be tuned according to your domain-specific task.
1.1 Performance Comparison
Developers can reproduce the performance comparison PR curves for TEN VAD and Silero VAD on the open-source testset (as shown in the figure above) by executing the following script on Linux x64 with a simply one line of code. The output figure will be saved in the same directory as the script.
cd ./examples
python plot_pr_curves.py
2. Agent-Friendly:
As illustrated in the figure below, TEN VAD rapidly detects speech-to-non-speech transitions, whereas Silero VAD suffers from a delay of several hundred milliseconds, resulting in increased end-to-end latency in human-agent interaction systems. In addition, as demonstrated in the 6.5s-7.0s audio segment, Silero VAD fails to identify short silent durations between adjacent speech segments.
3. Lightweight:
We evaluated the RTF (Real-Time Factor) across five distinct platforms, each equipped with varying CPUs. TEN VAD demonstrates much lower computational complexity and smaller library size than Silero VAD.
Platform
CPU
RTF
Lib Size
TEN VAD
Silero VAD
TEN VAD
Silero VAD
Linux
AMD Ryzen 9 5900X 12-Core
0.0150
/
306KB
2.16MB(JIT) / 2.22MB(ONNX)
Intel(R) Xeon(R) Platinum 8253
0.0136
Intel(R) Xeon(R) Gold 6348 CPU @ 2.60GHz
0.0086
0.0127
Windows
Intel i7-10710U
0.0150
/
464KB(x86) / 508KB(x64)
macOS
M1
0.0160
731KB
Android
Galaxy J6+ (32bit, 425)
0.0570
373KB(v7a) / 532KB(v8a)
Oppo A3s (450)
0.0490
iOS
iPhone6 (A8)
0.0210
320KB
iPhone8 (A11)
0.0050
4. Multiple programming languages and platforms:
TEN VAD provides cross-platform C compatibility across five operating systems (Linux x64, Windows, macOS, Android, iOS), with Python bindings optimized for Linux x64, with wasm for Web.
5. Supproted sampling rate and hop size:
TEN VAD operates on 16kHz audio input with configurable hop sizes (optimized frame configurations: 160/256 samples=10/16ms). Other sampling rates must be resampled to 16kHz.
Developers Testimonial
"We selected TEN VAD because it provides faster and more accurate sentence-end detection in Japanese compared to other VADs, while still being lightweight and fast enough for live use." - LiveCap,Hakase shojo.
"TEN VAD's overall performance is better than Silero VAD. Its high accuracy and low resource consumption helped us improve efficiency and significantly reduce costs." - Rustpbx.
Write your own use cases and import the class, the attributes of class TenVAD you can refer to ten_vad.py
from ten_vad import TenVad
JS Usage
1. Web
Requirements
Node.js (macOS v14.18.2, Linux v16.20.2 verified)
Terminal
Usage
1) cd ./examples
2) node test_node.js s0724-s0730.wav out.txt
C Usage
Build Scripts
Located in examples/ directory and examples_onnx (for ONNX usage on Linux):
Linux: build-and-deploy-linux.sh
Windows: build-and-deploy-windows.bat
macOS: build-and-deploy-mac.sh
Android: build-and-deploy-android.sh
iOS: build-and-deploy-ios.sh
Dynamic Library Configuration
Runtime library path configuration:
Linux/Android: LD_LIBRARY_PATH
macOS: DYLD_FRAMEWORK_PATH
Windows: DLL in executable directory or system PATH
Customization
Modify platform-specific build scripts
Adjust CMakeLists.txt
Configure toolchain and architecture settings
Overview of Usage
Navigate to examples/ or examples_onx/ (for ONNX usage on Linux)
Execute platform-specific build script
Configure dynamic library path
Run demo with sample audio s0724-s0730.wav
Processed results saved to out.txt
The detailed usage methods of each platform are as follows
1. Linux
Requirements
Clang (e.g. 6.0.0-1ubuntu2 verified)
CMake
Terminal
Note that if you did not install libc++1 (Linux), you have to run the code below to install it:
sudo apt update
sudo apt install libc++1
Usage (prebuilt-lib)
1) cd ./examples
2) ./build-and-deploy-linux.sh
Usage (ONNX)
You have to download the onnxruntime packages from the microsoft official onnxruntime github website. Note that the version of onnxruntime must be higher than or equal to 1.17.1 (e.g. onnxruntime-linux-x64-1.17.1.tgz).
You can check the official ONNX Runtime releases from this website. And for example, to download version 1.17.1 (Linux x64), use this link. After extracting the compressed file, you'll find two important directories:include/ - header files, lib/ - library files
1) cd examples_onnx/
2) ./build-and-deploy-linux.sh --ort-path /absolute/path/to/your/onnxruntime/root/dir
Note 1: If executing the onnx demo from a different directory than the one used when running build-and-deploy-linux.sh, ensure to create a symbolic link to src/onnx_model/ to prevent ONNX model file loading failures.
Note 2: The ONNX model locates in src/onnx_model directory.
2. Windows
Requirements
Visual Studio (2017, 2019, 2022 verified)
CMake (3.26.0-rc6 verified)
Terminal (MINGW64 or powershell)
Usage
1) cd ./examples
2) Configure "build-and-deploy-windows.bat" with your preferred:
- Architecture (default: x64)
- Visual Studio version (default: 2019)
3) ./build-and-deploy-windows.bat
3. macOS
Requirements
Xcode (15.2 verified)
CMake (3.19.2 verified)
Usage
1) cd ./examples
2) Configure "build-and-deploy-mac.sh" with your target architecture:
- Default: arm64 (Apple Silicon)
- Alternative: x86_64 (Intel)
3) ./build-and-deploy-mac.sh
4. Android
Requirements
NDK (r25b, macOS verified)
CMake (3.19.2, macOS verified)
adb (1.0.41, macOS verified)
Usage
1) cd ./examples
2) export ANDROID_NDK=/path/to/android-ndk # Replace it with your NDK installation path
3) Configure "build-and-deploy-android.sh" with your build settings:
- Architecture: arm64-v8a (default) or armeabi-v7a
- Toolchain: aarch64-linux-android-clang (default) or custom NDK toolchain
4) ./build-and-deploy-android.sh
Most questions can be answered by using DeepWiki, it is fast, intutive to use and supports multiple languages.
Citations
@misc{TEN VAD,
author = {TEN Team},
title = {TEN VAD: A Low-Latency, Lightweight and High-Performance Streaming Voice Activity Detector (VAD)},
year = {2025},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {https://github.com/TEN-framework/ten-vad.git},
email = {developer@ten.ai}
}
License
This project is licensed pursuant to the Apache 2.0 with additional conditions. Refer to the "LICENSE" file in the root directory for detailed information. Note that pitch_est.cc contains modified code derived from LPCNet, which is BSD-2-Clause and BSD-3-Clause licensed, refer to the NOTICES file in the root directory for detailed information.
