---
title: Austrian Dialect Text-to-Speech
emoji: 🗣️
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: 4.44.1
app_file: app.py
pinned: false
---

# Austrian Dialect Text-to-Speech Synthesis

A neural text-to-speech (TTS) interface demonstrating FastSpeech 2 with dialect embeddings for Austrian German dialects and standard German.

## Overview

This demo showcases neural speech synthesis for multiple Austrian dialects using:
- **FastSpeech 2** as the TTS backbone
- **IMS-Toucan** adaptation with dialect embeddings
- **VoxLingua107 XLS-R** for multilingual speaker embeddings
- **Standard German grapheme-to-phoneme conversion** with dialect support

Supported dialects:
- Standard Austrian German
- Viennese (Wien)
- Goisern (Upper Austrian)
- Innervillgraten (Tyrolean)

## Features

- **Multiple Speakers**: 12 different speaker profiles across all dialects
- **Dialect Selection**: Switch between standard and dialect embeddings
- **Interpolation**: Create custom dialect blends with 50/50 interpolations
- **German-only Interface**: Optimized for German text input

## Usage

1. Enter German text in the text input
2. Select a speaker from the dropdown
3. Choose a dialect/standard embedding
4. Optionally select an interpolation option
5. Click "Submit" to generate audio

## Examples

- "Der Nordwind und die Sonne stritten sich, wer von ihnen wohl der Stärkere wäre." (Standard)
- "Grüß Gott, wie geht es dir?" (Viennese)

## Technical Details

### Model Architecture
- **TTS Model**: FastSpeech 2
- **Vocoder**: Avocodo
- **Phonemizer**: espeak with Austrian dialect support
- **Embeddings**: XLS-R 300M wav2vec features for speaker/dialect conditioning

### Requirements
- Python 3.8+
- PyTorch 1.10.2
- Gradio 3.44.3
- All dependencies listed in `requirements.txt`

### File Structure
```
.
├── app.py                          # Main Gradio interface
├── requirements.txt                # Python dependencies
├── InferenceInterfaces/            # TTS inference implementations
├── Preprocessing/                  # Text preprocessing and embeddings
│   └── wav2vec_embeddings/        # Pre-computed dialect embeddings
├── Utility/                        # Utility functions
│   └── example_wavs/              # Speaker reference audio files
├── Models/                         # Pre-trained model files
└── audios/                         # Output audio directory
```

## References

[1] Y. Ren, C. Hu, X. Tan, T. Qin, S. Zhao, Z. Zhao, and T.-Y. Liu, "Fastspeech 2: Fast and high-quality end-to-end text to speech," in *ICLR 2021 - 9th International Conference on Learning Representations*, 2021.

[2] Florian Lux and Bilal Ghauri. IMS-Toucan: An Open-Source Framework for Training and Deploying Speech Synthesis Models. *Proc. INTERSPEECH 2023*, 2023.

[3] https://huggingface.co/TalTechNLP/voxlingua107-xls-r-300m-wav2vec

[4] L. Gutscher, M. Pucher, and V. García, "Neural Speech Synthesis for Austrian Dialects with Standard German Grapheme-to-Phoneme Conversion and Dialect Embeddings," in *Proc. 2nd Annual Meeting of the ELRA/ISCA SIG on Under-resourced Languages (SIGUL 2023)*, 2023. [Paper](https://www.isca-archive.org/sigul_2023/gutscher23_sigul.html)

## Authors

- [Lorenz Gutscher](mailto:lorenz.gutscher@ofai.at) - OFAI Vienna
- [Michael Pucher](mailto:michael.pucher@ofai.at) - OFAI Vienna

## Citation

If you use this work, please cite:

```bibtex
@inproceedings{gutscher2023neural,
  title={Neural Speech Synthesis for Austrian Dialects with Standard German Grapheme-to-Phoneme Conversion and Dialect Embeddings},
  author={Gutscher, Lorenz and Pucher, Michael and García, Víctor},
  booktitle={Proc. 2nd Annual Meeting of the ELRA/ISCA SIG on Under-resourced Languages (SIGUL 2023)},
  year={2023}
}
```

## License

Please refer to the LICENSE file for licensing information.

## Note on Data

Text input and audio output are processed temporarily during use. Please see the security notice in the interface for details.

---

For more information, visit the [Austrian Institute for Artificial Intelligence (OFAI)](https://www.ofai.at/)