---
license: apache-2.0
language:
  - km
  - cjm
tags:
  - text-to-speech
  - omnivoice
  - low-resource
---

# Multilingual TTS demo — 10 languages of Vietnam and Cambodia

A self-contained Gradio app. Clone the folder, install the requirements, run it.

```bash
pip install -r requirements.txt
python -u app.py
```

Everything resolves relative to `app.py`, so no paths need editing.

## Languages

| Code | Language | Code | Language |
|------|----------|------|----------|
| `km`  | Khmer    | `tyz` | Tay-Nung |
| `blt` | Tai Dam  | `ium` | Dao (Iu Mien) |
| `rad` | Ede      | `kpm` | Kho |
| `jra` | Jarai    | `cma` | Mnong |
| `bdq` | Bana     | `cjm` | Cham |

`blt` is **Tai Dam**, a Tai language of Vietnam written in Latin script — not
Thailand Thai, despite what the source dataset name suggests.

## Contents

```
app.py                 Gradio app
demo_voices/           one male + one female reference clip per language
checkpoint-60000/      model weights (inference only)
```

`demo_voices/voices.json` stores paths relative to itself, and each clip carries
the transcript needed as a voice-cloning reference. The clips come from a dev split
held out of training.

## Notes

- **Voice cloning** is the default: the chosen reference clip's speaker is copied.
  Upload your own clip plus its transcript, or pick *No reference* to let the model
  invent a voice for the language.
- **Long text** is split on sentence boundaries (including the Khmer khan `។`) and
  synthesized in batches, then joined. With no reference, the first segment becomes
  the reference for the rest so the speaker does not drift mid-passage.
- The checkpoint holds inference weights only. Optimizer/scheduler state for
  resuming training is not included.
- `app.py` registers its language ids with OmniVoice at startup. Without that step
  `generate()` silently drops any code outside its built-in table — nine of these
  ten are not in it — and would generate with no language conditioning at all.

## Model

OmniVoice (Qwen3-0.6B backbone + Higgs audio tokens) finetuned jointly on ~1,820 h
across the ten languages above, 60k steps.