DeepTalk-ASD

LR-ASD is a SOTA Active Speaker Detection (ASD) model. Building upon LR-ASD, this project provides an industrial-grade ASD system featuring full-pipeline ONNX conversion, integrated speaker verification, and real-time processing support. It runs efficiently on CPUs without relying on large PyTorch/GPU environments.

DeepTalk-ASD is an efficient Active Speaker Detection (ASD) system. By fusing audio, video features, and speaker embeddings, it determines in real-time which face in a video frame is speaking.

Key Features

• Multimodal Fusion: Deeply integrates face detection, voice activity detection (VAD), audio-visual ASD models, and Speaker Embeddings.
• Speaker Verification Enhancement (New):
  • Feature Fusion: Weighted fusion of ASD probability and audio speaker similarity to effectively suppress off-screen voice interference.
  • Fast Matching: Supports accelerated determination for tracks with known speaker embeddings, skipping complex calculations.
  • Automatic Profile Update: Automatically updates speaker embedding profiles for detected speakers using EMA (Exponential Moving Average).
• Modular Design:
  • FaceDetector: Provides face detection and tracking (currently supports InspireFace).
  • TurnDetector: Audio VAD and turn management (integrates Silero VAD).
  • SpeakerDetector: Core decision layer performing audio-visual feature extraction, fusion decision, and speaker comparison (based on LR-ASD ONNX).
• High Performance: Full-pipeline ONNX inference optimized for CPU, supporting real-time operation on mobile devices or standard laptops.
• Auto Model Management: Models are automatically downloaded on first use; supports offline mode and custom cache directories.
• Multiple Scenarios: Provides demos for real-time camera, video files, speaker embedding extraction, and pVAD.

System Architecture

The system works through three main sub-components:

1. Face Detection (FaceDetector): Locates and tracks faces (Tracks) in each video frame.
2. Turn Detection (TurnDetector): Determines if the current audio stream contains speech segments and manages their lifecycle (START, CONTINUE, END).
3. Speaker Detection (SpeakerDetector):
   • Extracts audio MFCC features and 112x112 grayscale mouth images.
   • Uses Sherpa-ONNX to extract speaker embeddings (Voiceprint).
   • Calculates raw ASD scores through a three-stage ONNX model (Audio/Visual Frontend + AV Backend).
   • Final scores are dynamically fused from ASD scores and speaker matching scores.

Quick Start

1. Install

pip install deeptalk_asd

2. Zero-Configuration Usage

Models are automatically downloaded on first use (~46 MB total). No manual setup required:

from deeptalk_asd import ASDDetectorFactory

asd = ASDDetectorFactory().create()

3. Running Demos

Core ASD Demos

• Real-time Camera Demo (Recommended first choice):

  python3 demo/realtime_asd_demo.py
  

• Offline Video File Processing:

  python3 demo/video_asd_demo.py --input demo/demo.mp4 --display
  

Model Management

Auto-Download (Default)

Models are downloaded to ~/.cache/deeptalk_asd/ on first use. Subsequent runs load from cache instantly.

Pre-Download Models (for offline environments)

# Download all models
python3 -m deeptalk_asd download-models

# Download to a specific directory
python3 -m deeptalk_asd download-models --cache-dir /path/to/models

# Check cache status
python3 -m deeptalk_asd info

Offline Mode

Set DEEPTALK_ASD_OFFLINE=1 to disable all network requests. Models must be pre-downloaded:

export DEEPTALK_ASD_OFFLINE=1
export DEEPTALK_ASD_CACHE_DIR=/path/to/models

Environment Variables

Configuration

Control components and their parameters precisely through the factory method:

from deeptalk_asd import ASDDetectorFactory

# Zero-configuration (recommended)
asd = ASDDetectorFactory().create()

# Custom configuration
config = {
    "face_detector": {
        "type": "inspireface",
        "model_dir": "weights"
    },
    "turn_detector": {
        "type": "silero-vad", 
        "model_dir": "weights"
    },
    "speaker_detector": {
        "type": "LR-ASD-ONNX", 
        "model_dir": "weights",
        "voiceprint_model_name": "wespeaker_zh_cnceleb_resnet34.onnx"
    }
}

asd = ASDDetectorFactory(**config).create()

License

The code in this project is licensed under the MIT License. However, the integrated pre-trained models are subject to their respective licenses:

1. InspireFace: Code is open-source, but weight files are typically restricted to non-commercial research. Please check official terms for commercial use.
2. Silero VAD: Licensed under the MIT License.
3. LR-ASD: Licensed under the MIT License.
4. Sherpa-ONNX: Licensed under the Apache-2.0 License.

Note: When releasing or commercializing applications based on this project, you must verify the license compliance for each model's weights.