Skip to content

Latest commit

 

History

History
155 lines (104 loc) · 5.39 KB

vad-models-setup.md

File metadata and controls

155 lines (104 loc) · 5.39 KB

VAD Models Setup Guide

This guide covers how to configure Voice Activity Detection models using the Editor UI provided by Unity-Sherpa-ONNX.

For runtime usage examples, see VAD Runtime Usage.

VAD Model Import

Opening the Settings

Project Settings > Sherpa-ONNX > VAD

The VAD settings window has three areas:

  • Import section — download a VAD model file by URL
  • Profile list — manage multiple VAD profiles
  • Profile detail — configure model paths, thresholds, and runtime options

Supported Model Types

Type Window Size Description
SileroVad 512 samples Silero VAD — compact and widely used
TenVad 256 samples TEN-VAD — alternative lightweight model

Model type is auto-detected from the file name during import.

Auto-Detection Keywords

File name keyword Detected type
silero SileroVad
ten TenVad

Importing a Model

VAD models are distributed as single .onnx files (not archives).

  1. Click Import from URL to expand the import section
  2. Paste the model file URL (direct link to .onnx file)
  3. Click Import

The importer downloads the file to Assets/StreamingAssets/SherpaOnnx/vad-models/{name}/, creates a profile, and auto-configures the model path.

Download URLs

Pre-trained models are available from the sherpa-onnx releases:

Silero VAD:

https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/silero_vad.onnx

TEN-VAD:

https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/ten-vad.onnx

For more information, see the sherpa-onnx VAD documentation.

Include VAD in Build

The Include VAD in Build toggle at the top of the settings controls whether VAD model files and settings are included in the build output:

  • Enabled (default) — VAD settings JSON and vad-models/ directory are included
  • Disabled — VAD files are excluded from the build, reducing app size

This is useful when VAD is only needed during development or when you want to ship separate builds with and without VAD support.

Profile Management

Creating a Profile

Click the + button below the profile list. A new profile is added and selected.

Removing a Profile

Select a profile and click the - button. The profile and its model directory are deleted.

Active Profile

Use the Active profile dropdown above the list to select which profile the runtime VAD system will use. This value is serialized to vad-settings.json at build time.

Profile Detail Fields

Identity

Field Description
Profile name Display name; also used as the model folder name
Model type SileroVad or TenVad

Thresholds

Field Default Description
Threshold 0.5 Speech probability threshold (0.0–1.0). Lower values detect more speech
Min silence duration 0.5 Minimum silence duration (seconds) to end a speech segment
Min speech duration 0.25 Minimum speech duration (seconds) to count as valid speech
Max speech duration 5.0 Maximum speech segment length (seconds); longer speech is split

Runtime

Field Default Description
Sample rate 16000 Expected input audio sample rate in Hz
Window size 512 Samples per detection window (512 for Silero, 256 for TEN-VAD)
Threads 1 Number of inference threads
Provider cpu ONNX Runtime execution provider
Buffer size (seconds) 60 Internal circular buffer capacity

Model

Field Description
Model Path to the .onnx model file (e.g. silero_vad.onnx)

Auto-Configure

If a model directory exists, the Auto-configure paths button appears at the top of the detail panel. Clicking it scans the model folder and fills the model path automatically by finding the first .onnx file.

Window Size and Model Type

When changing the model type in the profile detail, the window size is automatically adjusted:

  • SileroVad → 512 samples
  • TenVad → 256 samples

Using the wrong window size for a model type will produce incorrect results.

File Structure

Assets/StreamingAssets/SherpaOnnx/
  vad-settings.json            # Serialized profiles and active index
  vad-models/
    {profileName}/             # Model file for each profile
      silero_vad.onnx

Troubleshooting

Issue Solution
"Auto-configure paths" button missing Import or manually place the .onnx file in the profile's model directory
Import fails Ensure the URL points directly to an .onnx file, not an archive
Wrong model type detected Change the model type manually in the profile detail panel
Engine fails to load at runtime Verify the model path in vad-settings.json matches the actual file location
Detection seems off Adjust the threshold: lower values (e.g. 0.3) are more sensitive, higher values (e.g. 0.7) are stricter
Segments too short or fragmented Increase minSilenceDuration to merge nearby speech segments
Segments too long Decrease maxSpeechDuration to split long utterances
Wrong window size Ensure SileroVad uses 512 and TenVad uses 256; auto-adjusts on model type change