Skip to content
~/projects/avca
All projects
shippedGenerative Audio AI2025

Advanced Voice Conversion App

Voice-identity transformation using Wav2Vec2 and neural encoders, turn one speaker's voice into another while keeping the words.

Source
problem

The problem

Voice conversion has to disentangle *what* is said from *who* says it, then re-synthesize the content in a new identity without artifacts. Getting the content/speaker separation right is the whole game.

approach

How I approached it

  • Implemented voice-identity transformation with a Wav2Vec2-based feature pipeline and neural encoders.
  • Designed the preprocessing and inference workflows end to end.
  • Evaluated perceptual output quality to tune the pipeline.
architecture

Architecture

1Preprocess

Resample, trim, normalize source audio

2Encode

Wav2Vec2 features → content representation

3Convert

Neural encoder maps to target speaker identity

4Synthesize

Reconstruct waveform, perceptual QA

decisions

Engineering decisions & tradeoffs

$

Wav2Vec2 as the feature backbone

// Self-supervised speech features generalize far better than hand-built acoustic features for capturing linguistic content.

$

Perceptual evaluation, not just loss curves

// Voice quality is judged by ears; low reconstruction loss doesn't guarantee it sounds convincing.

stack
PythonPyTorchWav2Vec2LibrosaNumPy
what's next
  • › Real-time streaming conversion.
  • › Speaker embedding gallery for zero-shot target voices.
All projectsDiscuss this work