Fine-tuning and adaptation status

Inflect-Nano-v2 is an open-weight, inference-first release. The repository does not ship the private corpus-construction recipe, teacher setup, discriminator checkpoints, optimizer state, or a supported training command.

What is supported

Local inference from the released generator checkpoint.
Deterministic regression testing through a fixed text, seed, speed, and variation value.
Inspection of the complete deployable architecture and frontend.
Research adaptation in a separate VITS-compatible training environment, at the user's own risk.

What is not yet supported

A one-command new-voice trainer.
Zero-shot voice cloning or speaker embeddings.
Language adaptation without rebuilding the symbol inventory and frontend.
Resuming the release checkpoint as if it contained discriminators or optimizer state.

Data contract for future adapters

A defensible voice-adaptation toolkit must require a UTF-8 manifest with one audio path and one verified transcript per row; mono, consistently sampled, unclipped speech; held-out prompts that never enter training; speaker-identity and transcript audits; and an inference-only export that removes all training-only state. Duration, silence, clipping, and ASR checks must run before training rather than after a failed model.

Evaluation contract

Any adapted checkpoint should be compared with the frozen release on identical unseen prompts, a fixed seed policy, semantic WER, signal diagnostics, a learned quality predictor, and a blinded human listening test. Improving speaker similarity while reducing intelligibility is not a successful adaptation.

This file intentionally does not pretend that an unvalidated training recipe is a supported feature. A public adapter will be documented separately if it passes the same package and evaluation gates as the release models.