a2f-client

Audio2Face Client & Streaming Service

A comprehensive Python package that provides both a client library and a FastAPI streaming server for NVIDIA Omniverse Audio2Face (A2F). Generate facial blendshapes from audio with support for emotions, real-time streaming, and robust production-ready features.

---

🚀 Overview

This repository contains two main components:

1. a2f_client - A Python client library for programmatic control of Audio2Face via HTTP API
2. app - A FastAPI streaming server that processes audio files and streams blendshape data in real-time chunks

Key Use Cases:

• Real-time facial animation for avatars and VR/AR applications
• Batch processing of audio files for animation pipelines
• ML research and experimentation with facial expressions
• Integration into larger media production workflows

---

✨ Key Features

Client Library (a2f_client)

• Simple Python API for Audio2Face automation
• Emotion Control with customizable intensity weights
• Chunked Processing for large audio files
• Automatic USD Model Loading with ARKit blendshapes
• Robust Error Handling and logging
• Comprehensive Testing with pytest suite

Streaming Server (app)

• Real-time Streaming via FastAPI with JSON-line responses
• Dynamic FPS Optimization based on processing performance
• Parallel Processing using multiple A2F client instances
• Emotion Support for expressive facial animation
• Production Ready with logging, health checks, and error handling
• Multi-worker Support with file lock management

---

📁 Repository Structure

.
├── a2f_client/                 # Python client library
│   ├── modules/                # Internal HTTP API wrappers
│   ├── assets/                 # Default USD facial models
│   ├── samples/                # Example audio files
│   ├── tests/                  # Client unit tests
│   └── ...
├── app/                        # FastAPI streaming server
│   ├── config.py               # Server configuration
│   ├── main.py                 # FastAPI application
│   ├── streaming_logic.py      # Core streaming logic
│   └── tests/                  # Server integration tests
├── pyproject.toml              # Poetry configuration
└── README.md

---

🛠 Installation

Prerequisites

• Python 3.10+
• NVIDIA Omniverse Audio2Face (locally installed)
• Poetry (recommended) or pip

Setup

# Clone the repository
git clone git@github.com:alexthillen/a2f-client.git
cd a2f-client

# Install with Poetry
poetry install
poetry shell

# Or with pip
pip install -e .

Environment Configuration

Set your Audio2Face headless script path:

Definitions:

• Headless Script: Executable that launches Audio2Face in server mode without UI, required for API automation.

Linux:

export A2F_HEADLESS_SCRIPT="$HOME/.local/share/ov/pkg/audio2face-2023.2.0/audio2face_headless.sh"

Windows:

set A2F_HEADLESS_SCRIPT=%USERPROFILE%\AppData\Local\ov\pkg\audio2face-2023.2.0\audio2face_headless.bat

---

🎯 Quick Start

Python Client Usage

from a2f_client import A2FClient

# Initialize client
client = A2FClient(port=8192)

# Process audio with emotions
client.set_audio("path/to/audio.wav")
client.set_emotions({"joy": 0.8, "sadness": 0.2})

# Generate blendshapes for specific time window
blendshapes = client.generate_blendshapes(
    start=0.5, 
    end=1.0, 
    fps=30
)

print(f"Generated {blendshapes['blendshapes']['numFrames']} frames")

Streaming Server

Start the server:

python app/main.py
# Server starts at http://localhost:8000

Process audio via HTTP:

import requests
import json

# Simple audio processing
with open("audio.wav", "rb") as f:
    response = requests.post(
        "http://localhost:8000/process-audio/?fps=20",
        files={"audio_file": f},
        stream=True
    )
    
    for line in response.iter_lines():
        chunk_data = json.loads(line)
        print(f"Chunk {chunk_data['chunk_id']}: {chunk_data['result']['numFrames']} frames")

With emotion control:

emotions = {"joy": 0.8, "amazement": 0.3, "anger": 0.1}

with open("audio.wav", "rb") as f:
    response = requests.post(
        "http://localhost:8000/process-audio/",
        files={"audio_file": f},
        data={"emotions": json.dumps(emotions)},
        stream=True
    )

---

⚙️ Configuration

Client Settings (a2f_client/settings.py)

Server Settings (app/config.py)

---

🧪 Testing

Client Tests

# Run client unit tests
pytest a2f_client/tests/

# Test specific functionality
pytest a2f_client/tests/test_blendshapes.py::test_chunking_vs_full_export_equivalence

Server Tests

# Start server first
python app/main.py &

# Run integration tests
pytest app/tests/

# Test streaming with emotions
pytest app/tests/test_generation.py::test_streaming_endpoint_with_emotions

Test Coverage

• Client: Chunked vs full export probabilistic equivalence, emotion setting, audio loading
• Server: Streaming endpoints, dynamic FPS, emotion processing, error handling

---

🔧 Advanced Features

Dynamic FPS Optimization

The streaming server automatically adjusts frame rate based on processing performance:

# Algorithm used internally:
alpha = chunk_duration / processing_time
safe_fps = max(min_fps, floor(num_clients * current_fps * alpha - 7))
current_fps = min(max_fps, (current_fps + safe_fps) / 2)

This ensures optimal throughput while maintaining quality.

Emotion System

Supported Emotions:

• amazement, anger, cheekiness, disgust
• fear, grief, joy, outofbreath
• pain, sadness

Usage:

# Client
client.set_emotions({"joy": 1.0, "sadness": 0.3})

# Server API
emotions = {"joy": 0.8, "anger": 0.2}
requests.post(url, data={"emotions": json.dumps(emotions)}, ...)

Parallel Processing

The server uses multiple A2F client instances with file locking to prevent port conflicts:

# Automatic port allocation with locking
streaming_manager = StreamingManager(clients_per_worker=2)
# Handles 8190, 8191, 8192, etc. automatically

---

📊 API Reference

Client Methods

Definitions:

• Blendshape: Numerical weights representing facial muscle deformations for 3D animation
• USD Model: Universal Scene Description file containing 3D facial geometry and blendshape targets

Server Endpoints

---

🐛 Troubleshooting

Common Issues

1. "Headless script not found"

   • Set A2F_HEADLESS_SCRIPT environment variable to correct path
   • Verify Audio2Face installation

2. Port conflicts

   • Server uses ports 8190+ for A2F clients
   • Check for running A2F instances: netstat -tlnp | grep 819

3. File lock errors

   • Remove stale lock files: rm .lock_*
   • Ensure proper server shutdown

4. Audio format issues

   • Use WAV format for best compatibility
   • Check sample rate (44.1kHz recommended)

Performance Optimization

For High Throughput:

export A2F_FASTAPI_DEFAULT_WORKERS=3
export A2F_FASTAPI_CLIENTS_PER_WORKER=3
# Uses 9 A2F client instances total

For Low Latency:

streaming_manager.chunk_size = 0.1  # Smaller chunks

---

🔬 Important Terms & Theorems

Definitions:

• Blendshape: A vector of weights $∈ [1]^n$ representing facial expressions, where n is the number of facial control points
• Chunking: Temporal decomposition of audio into overlapping or non-overlapping segments for parallel processing
• A2F Instance: Individual Audio2Face headless server process bound to a specific port
• Emotion Weight: Continuous scalar $∈ [1]$ controlling the intensity of named emotional expressions

Chunk Equivalence Definition: Chunked blendshape generation is $(1%, 0.1)$-chunk-equivalent to full export generation if less than $1%$ of all blendshape weights disagree by more than an absolute value of $0.1$ or by more than $0.1 \times 10^{-6}$ relative error.

Implementation verified in test_chunking_vs_full_export_equivalence

---

📈 Summary Table

---

Fun Fact: The dynamic FPS optimization can automatically adjust from 10 FPS to 30 FPS based on your hardware's processing capability! 🚀