Amphion
An open-source audio generation toolkit supports speech synthesis, voice conversion and singing research, with MIT-licensed Python code.
Amphion is an audio generation toolkit for researchers and engineers developing speech and singing systems. It brings model implementations, dataset preparation and evaluation into one research codebase.
- Speech synthesis: Work with models including FastSpeech2, VITS, VALL-E, NaturalSpeech2 and MaskGCT, plus Vevo-TTS for control over voice timbre and style.
- Voice and accent conversion: Vevo changes voice identity or style, Noro supports noisy reference speech, and Vevo-Style handles zero-shot accent conversion.
- Singing generation: Vevo2 supports singing synthesis and conversion, voice editing, singing style changes and melody control.
- Text-to-audio: Generate audio from text with a latent diffusion model, and work with neural audio codecs including DualCodec and FACodec. Vocoders include HiFi-GAN and BigVGAN.
- Research evaluation: Compare generated audio for pitch, intelligibility, distortion and speaker similarity. Model visualizations help explain architectures, while shared preprocessing tools cover datasets such as AudioCaps, LibriTTS and LJSpeech.
Amphion is a Python toolkit with an MIT open-source license. Emilia-Pipe provides a preprocessing pipeline for speech data collected in real-world conditions.