> For the complete documentation index, see [llms.txt](https://langtech-bsc.gitbook.io/applied-intelligence/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://langtech-bsc.gitbook.io/applied-intelligence/models/speech-models/matxa-tts-central.md).

# matxa-tts-v2-ca-central-graphemes

Fast, natural Central Catalan text-to-speech that reads raw text directly, with no phonemizer needed.

<a href="https://huggingface.co/BSC-LT/matxa-tts-v2-ca-central-graphemes" class="button primary" data-icon="arrow-up-right-from-square">View on Hugging Face</a> <a href="https://colab.research.google.com/drive/14JzJHfFiEBwp1VwH4ZCuHlNSy2YgdPve" class="button secondary" data-icon="play">Try it in Colab</a>

**matxa-tts-v2-ca-central-graphemes** is a multi-speaker Catalan text-to-speech model specialized in **Central Catalan**. It works directly on **graphemes** (the written characters), so you don't need a phonemizer to turn text into phonemes first. That makes it easier to deploy than phoneme-based TTS.

It belongs to the second generation of **Matxa-TTS**, BSC's TTS model based on [Matcha-TTS](https://github.com/shivammehta25/Matcha-TTS). Matcha-TTS is a non-autoregressive architecture trained with conditional flow matching, which generates speech in a handful of steps.

## Key facts

|                  |                                                                                              |
| ---------------- | -------------------------------------------------------------------------------------------- |
| **Task**         | Text-to-speech (acoustic model: text → mel spectrogram)                                      |
| **Architecture** | Matcha-TTS v2 (OT-CFM flow matching, 1D U-Net decoder)                                       |
| **Input**        | Catalan text (graphemes)                                                                     |
| **Voices**       | Central Catalan voices from the Festcat and OpenSLR69 datasets                               |
| **Training**     | Pre-trained on 36 h of Common Voice across 5 accents, then fine-tuned on Festcat + OpenSLR69 |
| **Files**        | PyTorch checkpoint (213 MB) and an end-to-end ONNX model with a built-in vocoder (271 MB)    |
| **License**      | GPL-3.0                                                                                      |

## Evaluation

Evaluated on 350 sentences from Josep Carner's *La creació d'Eva i altres contes*. Quality is measured with ScoreQ, and intelligibility as Whisper's WER/CER on the synthesized audio:

| Model                                        | ScoreQ ↑ | WER ↓    | CER ↓    |
| -------------------------------------------- | -------- | -------- | -------- |
| **matxa-tts-v2-ca-central-graphemes**        | **3.32** | **6.84** | **2.09** |
| Matxa-TTS v1 multispeaker (previous version) | 2.96     | 9.03     | 3.40     |

## Quickstart

The quickest way to hear it is the [Colab notebook](https://colab.research.google.com/drive/14JzJHfFiEBwp1VwH4ZCuHlNSy2YgdPve), which runs the ONNX model with a vocoder. To run it locally, install BSC's Matcha-TTS fork:

```bash
git clone -b v2-graphemes --single-branch https://github.com/langtech-bsc/Matcha-TTS.git
cd Matcha-TTS
pip install -e .
```

{% hint style="warning" %}
**Normalize numbers first.** Input text containing digits must be normalized into words ("12" → "dotze") before synthesis.
{% endhint %}

### Choosing a vocoder

The model outputs mel spectrograms. A vocoder turns them into audio, and both of these work:

| Vocoder                                                             | License      | Note                                                                |
| ------------------------------------------------------------------- | ------------ | ------------------------------------------------------------------- |
| [alVoCat](https://huggingface.co/projecte-aina/alvocat-vocos-22khz) | CC BY-NC 4.0 | Catalan-trained vocoder used in the evaluation. **Non-commercial.** |
| [WaveNext](https://huggingface.co/BSC-LT/wavenext-mel)              | Apache 2.0   | Use this one for commercial deployments.                            |

## Good to know

* Quality can vary slightly from speaker to speaker.
* For other Catalan accents (Balearic, Valencian, North-Western, Northern), see [matxa-tts-v2-ca-multiaccent-graphemes](https://langtech-bsc.gitbook.io/applied-intelligence/models/speech-models/matxa-tts-multiaccent.md).

<details>

<summary>Funding and citation</summary>

Funded by the Ministerio para la Transformación Digital y de la Función Pública and by the EU (NextGenerationEU), within the framework of the project ILENIA (ref. 2022/TL22/00215337), and by the Generalitat de Catalunya through the Aina project. Training ran on MareNostrum 5.

```bibtex
@misc{LangtechVeu2025matxattsv2multispeaker,
  title={Catalan Multispeaker Text-to-Speech: Matxa-TTS v2},
  author={Jose Giraldo and Alex Peiró-Lilja},
  organization={Barcelona Supercomputing Center},
  url={https://huggingface.co/BSC-LT/matxa-tts-v2-ca-central-graphemes},
  year={2025}
}
```

</details>


---

# Agent Instructions
This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com.

## Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation by asking a question.

Perform an HTTP GET request on the following URL with the `ask` and `goal` query parameters:

```
GET https://langtech-bsc.gitbook.io/applied-intelligence/models/speech-models/matxa-tts-central.md?ask=<question>&goal=<user_goal>
```

`ask` is the immediate question: it should be specific, self-contained, and written in natural language.
`goal` is what the user is ultimately trying to achieve, the reason they need the answer. Sharing it helps GitBook give you a better, more relevant answer. A goal is most helpful when it describes the outcome the user wants rather than restating the question. For example, with `ask=how do I create an API token`, a goal like `automate deployments from our CI pipeline` lets GitBook tailor the answer to that use case.

The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.
