Skip to content

AUDIO_ANONYMIZATION.md

Latest commit

 

History

History
58 lines (42 loc) · 3.07 KB

AUDIO_ANONYMIZATION.md

File metadata and controls

58 lines (42 loc) · 3.07 KB
Raw

Audio Anonymization Features

This TTS router includes comprehensive audio anonymization to make it harder to identify distinguishing characteristics in the generated audio.

Features Implemented

1. Format Standardization

  • Convert all audio to MP3: Regardless of the original format from TTS providers
  • Consistent sample rate: All audio resampled to 44kHz
  • Mono conversion: Convert to single channel to reduce identifying characteristics
  • Consistent bitrate: All audio encoded at 128kbps

2. Complete Audio Reconstruction

  • Raw PCM extraction: Convert to raw PCM samples, completely stripping all container metadata and headers
  • Clean MP3 reconstruction: Re-encode from raw samples with minimal, standardized headers
  • Two-stage processing: Streamlined pipeline for maximum compatibility and effectiveness

3. Deep Binary-Level Cleaning

  • MP3 structure analysis: Analyzes MP3 file structure to identify audio frames vs metadata
  • Automatic text detection: Detects and neutralizes suspicious text sequences (4+ consecutive printable characters)
  • Metadata header removal: Identifies and removes binary signatures that could indicate metadata headers
  • Comprehensive sanitization: Neutralizes any non-audio data while preserving MP3 frame integrity

4. Generalized Approach

Instead of relying on specific signature lists, the system uses intelligent detection:

  • Text pattern analysis: Automatically identifies and removes text-like sequences
  • Binary signature detection: Recognizes metadata headers by structure rather than specific strings
  • Container-agnostic processing: Works regardless of the original audio format or embedded metadata
  • Future-proof: Effective against new or unknown identifying markers

Implementation Details

AudioProcessor Class

Located in tts_providers/audio_processor.py, this class handles all audio anonymization:

  • process_audio() - Main processing function for raw audio bytes
  • process_base64_audio() - Wrapper for base64-encoded audio (used by TTS providers)
  • _deep_clean_binary() - Advanced binary-level cleaning that analyzes MP3 structure and removes identifying data

Integration

The audio processor is automatically applied to all TTS provider outputs through the synthesize_speech() function in tts_providers/base.py.

Benefits for Anonymization

  1. Consistent output format: All audio has identical technical characteristics
  2. No provider fingerprinting: Removes TTS provider-specific metadata
  3. No encoder fingerprinting: Removes encoder-specific signatures
  4. Standardized quality: Consistent bitrate and sample rate across all outputs
  5. Minimal identifying information: Only contains the actual audio data

Dependencies

  • ffmpeg-python: Python bindings for FFmpeg audio processing
  • System FFmpeg installation with libmp3lame support

Usage

The audio anonymization is applied automatically to all TTS requests. No additional configuration is required - all audio returned by the /tts endpoint will be processed through the anonymization pipeline.