---
license: mit
language:
- fa
pipeline_tag: text-to-speech
datasets:
- MahtaFetrat/Mana-TTS
base_model:
- MahtaFetrat/Mana-Persian-Piper
tags:
- piper
- onnx
- persian
---

# DIO Piper FA

Pinned Persian Piper model used by DIO Voice. This release redistributes the
accepted Mana Persian Piper ONNX artifact unchanged; it is not the rejected
DIO prosody experiment.

## Files

- `fa_IR-dio-piper-fa.onnx`: Piper medium ONNX model.
- `fa_IR-dio-piper-fa.onnx.json`: Piper voice configuration.
- `sample-fa.mp3`: generated Persian sample.
- `SHA256SUMS`: release integrity hashes.

## Provenance

- Source model: `MahtaFetrat/Mana-Persian-Piper`, commit
  `ad9dd8518bedf517bd7cbc9f63b8e5c844bf5bc0`.
- Source model license declared by its model card: MIT.
- Training dataset: `MahtaFetrat/Mana-TTS`, declared CC0-1.0.
- Architecture lineage: Piper / VITS.

The DIO runtime adds a separate context-aware Persian phonemization stage. Its
GPL source and exact patches are distributed with the `dio-voice` source
release; those components are not embedded in this model repository.

## Use

Load the ONNX file with a compatible Piper runtime and its adjacent JSON
configuration. DIO Voice pins this repository by immutable Hub commit and
SHA-256; end-user installation does not require Hugging Face CLI or login.

## Evaluation and limitations

The artifact passed DIO's synthesis/PCM integration smoke check at 22,050 Hz.
No MOS, WER, speaker-similarity, or demographic benchmark is claimed. Persian
pronunciation and prosody remain dependent on input normalization and
phonemization. See `METRICS.md` for the release gate and known limits.

The experimental tail-locked checkpoint was excluded because it failed its
own prosody separation gate (`+0.72 st` and `+5.62 st/s` question-vs-statement
tail separation). This repository therefore makes no claim that DIO trained
or improved the acoustic weights.

## Sample

`sample-fa.mp3` says:

> ورزش بانوان در نیمی از استان‌های کشور وجود ندارد

## Attribution

Model authors and dataset contributors retain their original attribution.
DIO Voice packages the pinned artifact for portable inference.