thangquang09/parler-tts-vietnamese-v1-stage2

Vietnamese Parler-TTS Voice Design

Vietnamese instruction-guided text-to-speech built on top of Parler-TTS.

This model generates Vietnamese speech from:

• text: the content to speak
• description: a natural-language voice instruction describing accent, gender, age, pitch, speed, loudness, or speaking style

Demo Idea

Example voice descriptions:

• Giọng nữ trẻ miền Bắc, nói chậm rãi, giọng cao
• Giọng nam trưởng thành miền Nam, nói nhanh và rất to
• Giọng nữ miền Trung, nói chậm rãi với âm lượng nhỏ, giọng trầm

Model Details

• Model name: thangquang09/parler-tts-vietnamese-v1-stage2
• Architecture: Parler-TTS
• Language: Vietnamese
• Use case: instruction TTS / voice design
• Model type: controllable text-to-speech
• Base family: Parler-TTS
• Adaptation: Vietnamese fine-tuning with Vietnamese voice descriptions

How to Use

Python

import torch
import soundfile as sf
from transformers import AutoTokenizer
from parler_tts import ParlerTTSForConditionalGeneration

repo_id = "thangquang09/parler-tts-vietnamese-v1-stage2"
device = "cuda:0" if torch.cuda.is_available() else "cpu"

model = ParlerTTSForConditionalGeneration.from_pretrained(repo_id).to(device)
tokenizer = AutoTokenizer.from_pretrained(repo_id, use_fast=False)

text = "Xin chào, hôm nay bạn có khoẻ không?"
description = "Giọng nữ trẻ miền Bắc, nói chậm rãi, giọng cao"

desc_tokens = tokenizer(description, return_tensors="pt").to(device)
text_tokens = tokenizer(text, return_tensors="pt").to(device)

with torch.no_grad():
    generation = model.generate(
        input_ids=desc_tokens.input_ids,
        attention_mask=desc_tokens.attention_mask,
        prompt_input_ids=text_tokens.input_ids,
        prompt_attention_mask=text_tokens.attention_mask,
        do_sample=True,
        temperature=1.0,
    )

audio = generation.cpu().float().numpy().squeeze()
sf.write("output.wav", audio, model.config.sampling_rate)

Recommended Inference Repo

For a cleaner inference workflow with CLI and Gradio app, use the GitHub repository:

GitHub: https://github.com/thangquang09/vietnamese-parlertts

That repository provides:

• api.py for CLI and Python inference
• app.py for Gradio web UI
• simpler loading from Hugging Face or local checkpoints

Example CLI from the GitHub repo:

python api.py \
  --text "Xin chào, hôm nay bạn có khoẻ không?" \
  --description "Giọng nữ trẻ miền Bắc, nói chậm rãi, giọng cao" \
  --hf-repo thangquang09/parler-tts-vietnamese-v1-stage2 \
  --output output.wav

Input Format

This model expects two text inputs:

• Speech text: the Vietnamese sentence to synthesize
• Voice description: the natural-language instruction describing how the voice should sound

Good descriptions usually mention some of:

• gender
• age
• accent or region
• speaking rate
• pitch
• loudness
• expressiveness or emotion

Intended Use

This model is intended for:

• Vietnamese TTS research
• controllable speech generation
• voice style prompting
• demo systems and rapid prototyping

Limitations

• Voice control is prompt-based, so instruction following may vary by prompt quality
• Some accents, ages, or speaking styles may be stronger than others
• Performance may differ across text length and prompting style
• This is not a voice cloning model

Training Notes

This is a Vietnamese adaptation of Parler-TTS released for inference and downstream use.

If you want the cleaned inference-first codebase, scripts, and app, please use the GitHub repo: https://github.com/thangquang09/vietnamese-parlertts

Citation

If you use this model, please cite the original Parler-TTS work:

@misc{lacombe-etal-2024-parler-tts,
  author = {Yoach Lacombe and Vaibhav Srivastav and Sanchit Gandhi},
  title = {Parler-TTS},
  year = {2024},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/huggingface/parler-tts}}
}

@misc{lyth2024natural,
  title={Natural language guidance of high-fidelity text-to-speech with synthetic annotations},
  author={Dan Lyth and Simon King},
  year={2024},
  eprint={2402.01912},
  archivePrefix={arXiv},
  primaryClass={cs.SD}
}

Acknowledgements

• Parler-TTS
• Hugging Face
• Vietnamese adaptation and inference packaging by thangquang09