---
license: apache-2.0

language:
- ca

tags:
- TTS
- audio
- synthesis
- VITS
- speech
- coqui.ai
- pytorch

datasets:
- mozilla-foundation/common_voice_12_0
- projecte-aina/festcat_trimmed_denoised
- projecte-aina/openslr-slr69-ca-trimmed-denoised

---

# Aina Project's Catalan multi-speaker text-to-speech model
## Model description

This model was trained from scratch using the [Coqui TTS](https://github.com/coqui-ai/TTS) toolkit on a combination of 3 datasets: 
[Festcat](http://festcat.talp.cat/devel.php), [OpenSLR69](http://openslr.org/69/) and [Common Voice v12](https://commonvoice.mozilla.org/ca). 
For the training, we used 487 hours of recordings from 255 speakers. 
We have trimmed and denoised the data which all except Common Voice can be found in a seperate dataset 
in [festcat_trimmed_denoised](https://huggingface.co/datasets/projecte-aina/festcat_trimmed_denoised) and [openslr69_trimmed_denoised](https://huggingface.co/datasets/projecte-aina/openslr-slr69-ca-trimmed-denoised).

A live inference demo can be found in our spaces, [here](https://huggingface.co/spaces/projecte-aina/tts-ca-coqui-vits-multispeaker).

The model needs our fork of [espeak-ng](https://github.com/projecte-aina/espeak-ng) to work correctly. For installation and deployment please consult the docker file of our [inference demo](https://huggingface.co/spaces/projecte-aina/tts-ca-coqui-vits-multispeaker/blob/main/Dockerfile).

## Intended uses and limitations

You can use this model to generate synthetic speech in Catalan with different voices.

## How to use
### Usage

Required libraries:

```bash
pip install git+https://github.com/coqui-ai/TTS@dev#egg=TTS
```

Synthesize a speech using python:

```bash
import tempfile
import gradio as gr
import numpy as np
import os
import json

from typing import Optional
from TTS.config import load_config
from TTS.utils.manage import ModelManager
from TTS.utils.synthesizer import Synthesizer

model_path = # Absolute path to the model checkpoint.pth
config_path = # Absolute path to the model config.json
speakers_file_path = # Absolute path to speakers.pth file

text = "Text to synthetize"
speaker_idx = "Speaker ID"

synthesizer = Synthesizer(
    model_path, config_path, speakers_file_path, None, None, None,
)
wavs = synthesizer.tts(text, speaker_idx)
```


## Training
### Training Procedure
### Data preparation
### Hyperparameter

The model is based on VITS proposed by [Kim et al](https://arxiv.org/abs/2106.06103). The following hyperparameters were set in the coqui framework.

| Hyperparameter                     | Value                            |
|------------------------------------|----------------------------------|
| Model                              | vits                             |
| Batch Size                         | 16                               |
| Eval Batch Size                    | 8                                |
| Mixed Precision                    | false                            |
| Window Length                      | 1024                             |
| Hop Length                         | 256                              |
| FTT size                           | 1024                             |
| Num Mels                           | 80                               |
| Phonemizer                         | espeak                           |
| Phoneme Lenguage                   | ca                               |
| Text Cleaners                      | multilingual_cleaners            |
| Formatter                          | vctk_old                         |
| Optimizer                          | adam                             |
| Adam betas                         | (0.8, 0.99)                      |
| Adam eps                           | 1e-09                            |
| Adam weight decay                  | 0.01                             |
| Learning Rate Gen                  | 0.0001                           |
| Lr. schedurer Gen                  | ExponentialLR                    |
| Lr. schedurer Gamma Gen            | 0.999875                         |
| Learning Rate Disc                 | 0.0001                           |
| Lr. schedurer Disc                 | ExponentialLR                    |
| Lr. schedurer Gamma Disc           | 0.999875                         |

The model was trained for 730962 steps.

## Additional information

### Author
Language Technologies Unit (LangTech) at the Barcelona Supercomputing Center

### Contact information
For further information, send an email to aina@bsc.es

### Copyright
Copyright (c) 2023 Language Technologies Unit (LangTech) at Barcelona Supercomputing Center 

### Licensing Information
[Apache License, Version 2.0](https://www.apache.org/licenses/LICENSE-2.0)

### Funding
This work was funded by [Departament de la Vicepresidència i de Polítiques Digitals i Territori de la Generalitat de Catalunya](https://politiquesdigitals.gencat.cat/ca/inici/index.html#googtrans(ca|en) within the framework of [Projecte AINA](https://politiquesdigitals.gencat.cat/ca/economia/catalonia-ai/aina).

The training of the model was possible thanks to the compute time given by Galician Supercomputing Center CESGA ([Centro de Supercomputación de Galicia](https://www.cesga.es/)).

## Disclaimer
<details>
<summary>Click to expand</summary>

The models published in this repository are intended for a generalist purpose and are available to third parties. These models may have bias and/or any other undesirable distortions.

When third parties, deploy or provide systems and/or services to other parties using any of these models (or using systems based on these models) or become users of the models, they should note that it is their responsibility to mitigate the risks arising from their use and, in any event, to comply with applicable regulations, including regulations regarding the use of Artificial Intelligence.

In no event shall the owner and creator of the models (BSC – Barcelona Supercomputing Center) be liable for any results arising from the use made by third parties of these models.