Inflect-Nano-v2 fine-tuning toolkit released

A new fine-tuning toolkit has been released that lets the ultra-compact TTS models Inflect-Nano-v2 (3.96M parameters) and Inflect-Micro-v2 (9.35M parameters) be trained on a user's own voice or on other languages.

Key features of this toolkit include:

• Custom training: Uses single-speaker recordings and transcripts to warm-start the model or resume training.
• Multilingual support: Handles new phoneme inventories, enabling adaptation to languages other than English.
• Flexible export: Trained results can be exported in PyTorch or ONNX format, allowing deployment across a variety of environments.
• Efficient architecture: A text-to-waveform model that includes a 24kHz waveform decoder, operating with very few parameters.

This toolkit is not zero-shot cloning, but rather a supervised adaptation approach optimized for a specific speaker and language.