---
title: Vietnamese TTS - SparkTTS
emoji: 🎙️
colorFrom: blue
colorTo: green
sdk: gradio
sdk_version: 5.9.1
app_file: app.py
pinned: false
license: mit
---

# Vietnamese TTS Demo

A Vietnamese text-to-speech demo powered by **SparkTTS** fine-tuned on Vietnamese speech data, using the [MiraTTS](https://github.com/ysharma3501/MiraTTS) inference pipeline.

## How it works

1. Upload a short reference audio clip (3–10 seconds) to clone the voice
2. Enter Vietnamese text
3. Click **Tổng hợp giọng nói** to generate speech

The model uses a BiCodec audio tokenizer to encode the reference voice, then an LLM generates speech tokens conditioned on the text, and a neural decoder + super-resolution upsampler produces 48kHz audio output.

## Connecting your private model

In your Space settings, add these secrets/variables:

| Key | Value |
|-----|-------|
| `HF_TOKEN` | A HuggingFace token with read access to the private model repo |
| `MODEL_ID` | (optional) Override model repo, default: `sleeper371/sparktts` |

## Architecture

- **LLM backbone**: SparkTTS-0.5B (LLaMA-based) fine-tuned on Vietnamese
- **Audio codec**: BiCodec (ONNX) from [FastBiCodec](https://github.com/ysharma3501/FastBiCodec)
- **Super-resolution**: [FlashSR](https://github.com/ysharma3501/FlashSR) → 48kHz output
- **Inference**: `transformers` (ZeroGPU-compatible, no lmdeploy)