Skip to main content
PORTFOLIO
BACK TO ALL PROJECTS
AGENTIC AIProduction Live

SystemAccentChanger — Real-Time AI Audio Accent Converter & Virtual Mic

Real-time bidirectional speech conversion pipeline integrating Google Gemini Live WebSockets with a PulseAudio/PipeWire virtual microphone null-sink and low-latency DSP noise gate.

PythonGoogle Gemini LivePulseAudioPipeWireWebSocketsDSPAudio Engineering
Source Code

🎙️ System Accent Changer - AI Virtual Microphone Studio

A real-time Linux audio engine and virtual microphone powered by Google Gemini Live Speech-to-Speech (STS). It captures your speech from your physical microphone, removes hardware noise and DC bias, streams audio to Gemini Live to convert accent, tone, and grammar in real time, and feeds the synthesized voice into a virtual system microphone.

Any application that uses a microphone (such as Google Meet, Zoom, Discord, Slack, or OBS) can select the virtual microphone to receive the converted audio.

🌟 Key Features

  • Real-Time Gemini Live STS: Bidirectional audio streaming over WebSocket with ultra-low latency using gemini-3.1-flash-live-preview.
  • Accent & Tone Rewriting: Convert regional accents (e.g. Indian, Chinese, Spanish) to target accents (American, British, Australian, etc.) with customizable registers (professional, friendly, casual, formal, empathetic).
  • Verbatim Relay Mode: Option to relay speech directly with clean native pronunciation without altering grammar or words.
  • Hardware DC Blocker & Voice Activity Detection (VAD): Intelligent energy-based noise gate with automatic DC offset filtering for laptop analog microphones.
  • Virtual PipeWire / PulseAudio Routing: Creates a virtual audio sink and remapped source compatible with all Linux desktop environments.
  • Auto-reconnection: Gracefully recovers from WebSocket idle timeouts or network interruptions.

🏗️ Architecture

[ Physical Microphone ] ──► (alsa_input / hw mic)
           │
           ▼
[ Hardware DC Blocker & NoiseGate ] ──► (Filters hum & DC offset)
           │
           ▼
[ Gemini Live STS Client ] ──► (WebSocket Bidirectional Audio)
           │                   - Transcribes input intent
           │                   - Rewrites tone / accent
           │                   - Synthesizes 24kHz audio
           ▼
[ Resampler (24kHz → 48kHz) ]
           │
           ▼
[ Virtual Null Sink ] ──────► ('US_Voice_Sink' or 'ai_mic')
           │
           ▼
[ Remapped Virtual Source ] ─► ('US_Voice_Mic' / 'AI_Virtual_Microphone')
           │
           ▼
[ Video Call / Apps ] ──────► (Zoom, Google Meet, Discord, Slack)

📋 Prerequisites

1. System Audio Libraries (Linux / Ubuntu / Debian)

This project uses sounddevice which relies on PortAudio:

sudo apt update
sudo apt install -y portaudio19-dev libasound2-dev pulseaudio-utils

2. Python Environment

Python 3.10 or newer (tested up to Python 3.14).

⚙️ Installation & Setup

Step 1: Create and Activate Virtual Environment

# Clone the repository
git clone git@github.com:kartikgetitnow-spec/SystemAccentChanger.git
cd SystemAccentChanger

# Create virtual environment
python3 -m venv .venv

# Activate virtual environment
source .venv/bin/activate

(On Windows: .venv\Scripts\activate)

Step 2: Install Dependencies

pip install --upgrade pip
pip install -r requirements.txt

🔐 Environment Configuration (.env)

[!WARNING] Never commit your .env file to Git! It contains your secret Google Gemini API key. The repository's .gitignore is pre-configured to ignore .env.

Create your private .env file from the provided template:

cp .env.example .env

Open .env in an editor and configure the parameters:

# ------------------------------------------------------------------------------
# 1. Gemini API Credentials
# ------------------------------------------------------------------------------
# Obtain your free API key at: https://aistudio.google.com/app/apikey
GEMINI_API_KEY=your_gemini_api_key_here

# Gemini Live bidirectional model
GEMINI_MODEL=gemini-3.1-flash-live-preview
GEMINI_LIVE_MODEL=gemini-3.1-flash-live-preview

# AI Voice Gender: male | female (default: male)
VOICE_GENDER=male

# Optional: manually override with a specific voice name (Puck, Charon, Aoede, Fenrir, Kore)
# GEMINI_VOICE=Puck

# ------------------------------------------------------------------------------
# 2. Audio Hardware & Virtual Devices
# ------------------------------------------------------------------------------
# Physical microphone name or substring.
# Setting 'alsa_input' binds directly to your real mic and prevents loopbacks.
INPUT_DEVICE=alsa_input
OUTPUT_DEVICE=default

# Virtual sink device where output is written
VIRTUAL_MIC_DEVICE=US_Voice_Sink

# ------------------------------------------------------------------------------
# 3. Stream & Format
# ------------------------------------------------------------------------------
SAMPLE_RATE=16000
VIRTUAL_MIC_SAMPLE_RATE=48000
CHANNELS=1
CHUNK_SIZE=2048

# ------------------------------------------------------------------------------
# 4. Accent & Tone Conversion Settings
# ------------------------------------------------------------------------------
ENABLE_ACCENT_CONVERSION=true
TARGET_ACCENT=american       # american | british | australian | indian | neutral
SOURCE_ACCENT=indian         # indian | chinese | spanish | french | etc.
ACCENT_MODE=professional     # professional | friendly | casual | formal | empathetic
TARGET_LANGUAGE=english

# ------------------------------------------------------------------------------
# 5. Noise Gate & Logging
# ------------------------------------------------------------------------------
ENABLE_NOISE_GATE=true
NOISE_GATE_THRESHOLD_DB=10.0
NOISE_GATE_HANGOVER_MS=300.0
LOG_LEVEL=INFO

🎛️ Setting Up the Virtual Microphone on Linux

To create the virtual audio devices in PulseAudio / PipeWire, run the setup script:

chmod +x scripts/setup_virtual_mic.sh
./scripts/setup_virtual_mic.sh

Or execute manually:

# 1. Create the null sink
pactl load-module module-null-sink sink_name=US_Voice_Sink sink_properties=device.description=US_Voice_Sink

# 2. Create the virtual microphone source
pactl load-module module-remap-source master=US_Voice_Sink.monitor source_name=US_Voice_Mic source_properties="device.description=US_Voice_Mic"

To verify the audio devices are active:

python main.py --list-devices

🚀 Running the Project

1. Dry-Run Mode (Test Without API)

Echoes your microphone directly to the virtual microphone without calling the Gemini API. Useful for testing latency and audio loopback:

python main.py --dry-run

2. Live AI Accent Conversion Mode

Starts the real-time AI conversion pipeline:

python main.py

3. CLI Overrides & Options

You can override any .env setting directly from the command line:

# Choose AI voice gender (male | female) - Default is male
python main.py --gender female
python main.py --gender male

# Aliases also supported:
python main.py --voice-type female
python main.py --voice-gender male

# Convert to British accent with female voice and friendly tone
python main.py --target-accent british --gender female --mode friendly

# Verbatim speech relay (keep original words, speak clearly with target accent)
python main.py --no-accent-conversion

# Use a specific Gemini voice directly by name (Puck, Charon, Aoede, Fenrir, Kore)
python main.py --voice Charon

# Pass a custom system instruction prompt
python main.py --system-prompt "You are a voice translator. Translate Spanish speech to clear English."

# View available audio devices and active voice configuration
python main.py --list-devices

🎧 Using in Apps (Zoom, Google Meet, Discord, etc.)

  1. Open Google Meet, Zoom, Discord, or system settings.
  2. In the Audio / Input Device dropdown, select:
    • US_Voice_Mic (or AI_Virtual_Microphone).
  3. Set your Output Device (Speakers/Headphones) to your physical headphones.
  4. Speak normally into your physical microphone — participants will hear your voice transformed in real time!

❓ Troubleshooting & FAQs

Why do I hear a double voice / echo?

  • Acoustic Feedback: If your speakers are on, your laptop microphone will pick up Gemini's converted voice and send it back to Gemini, causing a repeating echo. Always wear headphones when testing.
  • Mic Monitoring / Sidetone: Make sure your operating system or meeting app does not have "Listen to this device" or mic sidetone turned on.
  • Multiple Pipeline Instances: Ensure only one instance of main.py is running (pkill -f "python main.py").

Why does Gemini disconnect with Error 1008 (The operation was aborted)?

  • Gemini Live closes idle WebSocket sessions if no audio is received for ~2.5 minutes.
  • The pipeline automatically reconnects. To avoid long idle disconnects, ensure INPUT_DEVICE points to your real microphone (INPUT_DEVICE=alsa_input) so your speech is detected.

How do I stop the pipeline?

Press Ctrl + C in the terminal to trigger a clean shutdown of all audio streams.

📜 License

MIT License. Free for personal and commercial use.