File size: 3,087 Bytes
b9c74dd 3cf87cd b9c74dd 98aa56d b9c74dd 98aa56d b9c74dd |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 |
---
license: mit
language:
- lb
tags:
- text-to-speech
- tts
- vits2
- luxembourgish
pipeline_tag: text-to-speech
---
# VITS2 - Claude (Luxembourgish Gender-Neutral Voice)
A VITS2-based text-to-speech model for Luxembourgish, featuring a synthetic gender-neutral voice.
## Model Description
This model was trained using the VITS2 architecture on Luxembourgish speech data from the [Lëtzebuerger Online Dictionnaire (LOD)](https://lod.lu) example sentences.
"Claude" is a synthetic gender-neutral Luxembourgish voice created by modulating the original LOD recordings.
### Model Details
- **Architecture:** VITS2 with duration discriminator and transformer flows
- **Language:** Luxembourgish (lb)
- **Speaker:** Single speaker (gender-neutral, synthetic)
- **Sample Rate:** 24000 Hz
- **Checkpoint:** G_57000 (57,000 steps)
- **License:** MIT
## Usage
**Note:** Text should be lowercased before synthesis. Additional text normalization may be required.
This model requires the included Python source files for inference.
### Basic Usage
```python
import torch
import scipy.io.wavfile as wavfile
from vits2_engine import VITS2Engine
# Load the model
engine = VITS2Engine(model_dir="path/to/vits2-claude")
# Generate speech
wav = engine.tts("moien, wéi geet et dir?")
# Save to file
wavfile.write("output.wav", engine.sample_rate, wav)
```
### Command Line
```bash
python inference.py "moien, wéi geet et dir?"
# With custom parameters
python inference.py "Text" --noise_scale 0.5 --length_scale 1.1 -o output.wav
```
### Parameters
- `noise_scale`: Controls voice variation (default: 0.667, lower = more consistent)
- `noise_scale_w`: Controls duration variation (default: 0.8)
- `length_scale`: Controls speech speed (default: 1.0, higher = slower)
## Technical Specifications
| Parameter | Value |
|-----------|-------|
| Hidden Channels | 192 |
| Filter Channels | 768 |
| Attention Heads | 2 |
| Encoder Layers | 6 |
| Mel Channels | 80 |
| FFT Size | 1024 |
| Hop Length | 256 |
## Requirements
- Python 3.8+
- PyTorch
- scipy
- numpy
- Cython (for monotonic_align)
## Citation
If you use this model, please cite:
```bibtex
@misc{zls2025vits2claude,
title={VITS2 Claude - Luxembourgish Gender-Neutral Voice},
author={Zenter fir d'Lëtzebuerger Sprooch},
year={2025},
publisher={Hugging Face},
url={https://huggingface.co/ZLSCompLing/VITS2-Claude}
}
```
## Acknowledgments
Developed by [Zenter fir d'Lëtzebuerger Sprooch](https://zls.lu).
Voice data sourced from the [Lëtzebuerger Online Dictionnaire (LOD)](https://lod.lu). The original audio files are available via the [LOD linguistic data on data.public.lu](https://data.public.lu/en/datasets/letzebuerger-online-dictionnaire-lod-linguistesch-daten/), which provides an XML file containing example sentence IDs. Audio files can be accessed at:
```
https://lod.lu/uploads/examples/AAC/{folder}/{id}.m4a
```
where `{folder}` is the first 2 characters of `{id}`.
This model is used in [Sproochmaschinn](https://sproochmaschinn.lu), a Luxembourgish speech processing platform.
|