README.md

---
language:
- ca
licence:
- apache-2.0
tags:
- matcha-tts
- acoustic modelling
- speech
- multispeaker
pipeline_tag: text-to-speech
datasets:
- projecte-aina/festcat_trimmed_denoised
- projecte-aina/openslr-slr69-ca-trimmed-denoised
---

# Matcha-TTS Catalan Multispeaker

## Table of Contents
<details>
<summary>Click to expand</summary>

- [Model description](#model-description)
- [Intended uses and limitations](#intended-uses-and-limitations)
- [How to use](#how-to-use)
- [Training](#training)
- [Evaluation](#evaluation)
- [Citation](#citation)
- [Additional information](#additional-information)

</details>

## Model Description

**Matcha-TTS** is an encoder-decoder architecture designed for fast acoustic modelling in TTS. 
The encoder part is based on a text encoder and a phoneme duration prediction that together predict averaged acoustic features.
And the decoder has essentially a U-Net backbone inspired by [Grad-TTS](https://arxiv.org/pdf/2105.06337.pdf), which is based on the Transformer architecture. 
In the latter, by replacing 2D CNNs by 1D CNNs, a large reduction in memory consumption and fast synthesis is achieved.

**Matcha-TTS** is a non-autorregressive model trained with optimal-transport conditional flow matching (OT-CFM). 
This yields an ODE-based decoder capable of generating high output quality in fewer synthesis steps than models trained using score matching.

## Intended Uses and Limitations

This model is intended to serve as an acoustic feature generator for multispeaker text-to-speech systems for the Catalan language. 
It has been finetuned using a Catalan phonemizer, therefore if the model is used for other languages it may will not produce intelligible samples after mapping 
its output into a speech waveform. 

The quality of the samples can vary depending on the speaker. 
This may be due to the sensitivity of the model in learning specific frequencies and also due to the quality of samples for each speaker.

## How to Use

### Installation

This model has been trained using the espeak-ng open source text-to-speech software. 
The espeak-ng containing the Catalan phonemizer can be found [here](https://github.com/projecte-aina/espeak-ng)

Create a virtual environment:
```bash
python -m venv /path/to/venv
```
```bash
source /path/to/venv/bin/activate
```

For training and inferencing with Catalan Matcha-TTS you need to compile the provided espeak-ng with the Catalan phonemizer:
```bash
git clone https://github.com/projecte-aina/espeak-ng.git

export PYTHON=/path/to/env/<env_name>/bin/python
cd /path/to/espeak-ng
./autogen.sh
./configure --prefix=/path/to/espeak-ng
make
make install

pip cache purge
pip install mecab-python3
pip install unidic-lite
```
Clone the repository:

```bash
git clone -b dev-cat-hf https://github.com/langtech-bsc/Matcha-TTS.git

```
Install its requirements:
```bash
pip install requirements.txt

```


### For Inference

#### PyTorch

Speech inference can be done with **Catalan Matcha-TTS** by loading the model remotely with the HF hub.  

```bash
pip install git+https://github.com/langtech-bsc/Matcha-TTS.git@dev-cat

```

#### ONNX

We also release a ONNX version of the model

### For Training

The entire checkpoint is also release to continue with pretraining or finetuning. 

## Training Details

### Training data

The model was trained on 2 **Catalan** speech datasets

| Dataset             | Language | Hours   | Num. Speakers   |
|---------------------|----------|---------|-----------------|
| [Festcat](https://huggingface.co/datasets/projecte-aina/festcat_trimmed_denoised)             | ca       | 22      | 11              |
| [OpenSLR69](https://huggingface.co/datasets/projecte-aina/openslr-slr69-ca-trimmed-denoised)           | ca       | 5       | 36              |

### Training procedure

***Catalan Matcha-TTS*** was not trained from scratch. Instead, we finetuned the model from the English multispeaker checkpoint 
(trained with the [VCTK dataset](https://huggingface.co/datasets/vctk)) provided by the authors. 
The embedding layer was initialized with the number of catalan speakers (47) and original hyperparameters were kept.

### Training Hyperparameters

* batch size: 32 (x2 GPUs)
* learning rate: 1e-4
* number of speakers: 47
* n_fft: 1024
* n_feats: 80
* sample_rate: 22050
* hop_length: 256
* win_length: 1024
* f_min: 0
* f_max: 8000
* data_statistics:
  * mel_mean: -6578195
  * mel_std: 2.538758
* number of samples: 13340

## Evaluation

Validation values obtained from tensorboard from epoch 2399*: 

* val_dur_loss_epoch: 0.38
* val_prior_loss_epoch: 0.97
* val_diff_loss_epoch: 2.195

(Note that the finetuning started from epoch 1864, as previous ones were trained with VCTK dataset)

## Citation

If this code contributes to your research, please cite the work:

```
@misc{mehta2024matchatts,
      title={Matcha-TTS: A fast TTS architecture with conditional flow matching}, 
      author={Shivam Mehta and Ruibo Tu and Jonas Beskow and Éva Székely and Gustav Eje Henter},
      year={2024},
      eprint={2309.03199},
      archivePrefix={arXiv},
      primaryClass={eess.AS}
}
```

## Additional Information

### Author
The Language Technologies Unit from Barcelona Supercomputing Center.

### Contact
For further information, please send an email to <langtech@bsc.es>.

### Copyright
Copyright(c) 2023 by Language Technologies Unit, Barcelona Supercomputing Center.

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

### Funding
This work has been promoted and financed by the Generalitat de Catalunya through the [Aina project](https://projecteaina.cat/).