Voice and language adaptation

Inflect v2 is an open-weight, inference-first release. The public checkpoint is the complete deployable generator, not a resumable training snapshot: it does not contain optimizer state, discriminators, training manifests, or the private corpus-construction pipeline.

This document separates what is technically possible from what is supported.

Current support level

Task	Status	What that means
Local inference	Supported	Tested through the packaged Python API and CLI.
Continue generator training	Research use	The released generator can initialize a compatible VITS-family training stack.
Replace the fixed voice	Research use	Requires a new single-speaker corpus and careful decoder adaptation.
Add a selectable second voice	Not supported	The release architecture has no speaker-ID or speaker-embedding interface.
Add a language	Not supported	Requires a new frontend, symbol inventory, embeddings, alignments, and data.
Zero-shot cloning	Not supported	There is no reference encoder or cloning interface.

What a compatible adaptation stack needs

The release config describes the deployable generator:

24 kHz mono waveform output;
16,384-sample training segments;
English phoneme input with interspersed blanks;
stochastic latent synthesis and monotonic alignment;
an integrated waveform decoder.

A separately reconstructed trainer must add the training-only components that are intentionally absent from the release:

waveform and period discriminators compatible with the decoder;
optimizer and learning-rate schedule state;
mel, KL, duration, feature-matching, and adversarial losses;
data loading, segmentation, and alignment handling;
validation synthesis and checkpoint export.

Loading model.pth as the generator initialization is reasonable. Treating it as a complete resumable checkpoint is not.

Data contract

Use a UTF-8 manifest with one verified transcript and one audio path per row. The audio should be:

mono, consistently sampled, and free of clipping;
one speaker with stable recording conditions;
trimmed without cutting consonants or sentence endings;
diverse in phonemes, punctuation, sentence shape, numbers, and names;
split into train and held-out evaluation sets before training.

Run transcript verification, duration checks, silence checks, clipping checks, and speaker-consistency checks before training. Voice adaptation quality is usually limited by data coverage and decoder mismatch before it is limited by the number of optimization steps.

Conservative warm-start sequence

This is research guidance, not a guaranteed recipe:

Load the public generator and freshly initialize training-only components.
Begin with a low generator learning rate.
Temporarily freeze most of the waveform decoder while the text and latent paths adapt to the new speaker.
Unfreeze the decoder gradually and monitor high-frequency buzz, sibilance, identity drift, and intelligibility.
Export inference-only candidates frequently.
Select by held-out listening and matched evaluation, not training loss.

Do not publish an adapted voice as equivalent to the release voice without a matched unseen-text evaluation.

Required evaluation

Compare every candidate with the frozen release on identical prompts and seeds:

semantic WER with more than one ASR evaluator;
clipping, silence, loudness, and ending diagnostics;
a learned quality predictor used only as a diagnostic;
speaker consistency for voice replacement;
blinded human listening on unseen text.

An adaptation that improves speaker similarity while reducing intelligibility, stability, or audio quality is not a successful adaptation.

Why there is no one-command trainer

A nominal script would be easy to publish; a trustworthy trainer is not. The project will call adaptation supported only after a clean environment can produce a usable checkpoint, export it correctly, and pass the same package and evaluation gates as the release models. Until then, the honest interface is the tested inference package plus the architecture/configuration needed for independent research.