Contributing to Inflect

July 24, 2026 · View on GitHub

Contributions are welcome. For a speech model, the bar is reproducible evidence rather than a single favorable clip.

Good Contributions

Useful contributions usually fall into one of these categories:

reproducible benchmark improvements
clearer docs
safer data preparation
better evaluation prompts
runtime stability fixes
small, well-scoped runtime experiments
bug reports with audio examples and exact commands

Before Opening a PR

Please make sure:

the change is small enough to review
generated artifacts are not committed
private reference voices are not committed
checkpoints are not committed
local absolute paths are not added to public docs
the README does not claim unreleased model quality

Runtime and evaluation reports

For model or inference changes, include:

base variant
changed variant
exact command
prompt set
model package and commit tested
checkpoint used
listening notes
objective metrics if available

Minimum listening notes:

voice consistency
pacing
skipped words
glitches
long-prompt behavior

Commit Scope

Keep PRs focused.

Good:

"Add ASR pseudo-label filter"
"Improve README and media kit"
"Add duration-ratio check to benchmark"

Bad:

one PR containing docs, checkpoints, generated audio, unrelated training changes, and local state files

Local artifacts

Do not commit:

outputs/
.blind_ab_state*/
reference_voices/
local third-party checkouts
checkpoints
full generated datasets
private audio

The release is open-weight. A contribution may improve public inference, evaluation, examples, documentation, or integration without requiring the private corpus-construction pipeline. Do not open issues requesting private reference material, generated corpora, credentials, or undisclosed training infrastructure.

Project tone

Inflect should be ambitious without exaggerating. Distinguish measured results from estimates, human preference from predicted quality, and complete text-to-waveform parameters from training-only modules.