---
license: cc0-1.0
tags:
- text-to-speech
- tts
- speech-synthesis
- persian
- data-collection
- data-preprocessing
- speech-processing
- forced-alignment
- speech-dataset
- speech-corpus
- dataset-preparation
- persian-speech
- tts-dataset
- text-to-speech-dataset
- mana-tts
- manatts
- speech-data-collection
---

# ManaTTS Persian: a recipe for creating TTS datasets for lower resource languages

![Hugging Face](https://img.shields.io/badge/Hugging%20Face-Dataset-orange)

Mana-TTS is a comprehensive Persian Text-to-Speech (TTS) dataset, featuring 102 hours of high-quality single-speaker audio, specifically designed for speech synthesis and related tasks. The dataset has been carefully collected, processed, and annotated to ensure high-quality data for training TTS models. For details on data processing pipeline and statistics, please refer to the paper in the Citation section.

## Pretrained Models
A Tacotron2 model has been trained on this dataset and is available [here.](https://huggingface.co/MahtaFetrat/Persian-Tacotron2-on-ManaTTS)

## Acknowledgement
The raw audio and text files have been collected from the archive of [Nasl-e-Mana](https://naslemana.com/) magazine devoted to the blind.
We thank the Nasl-e-Mana magazine for their invaluable work and for being so generous with
the published dataset license. We also extend our gratitude to the [Iran Blind Non-governmental
Organization](https://ibngo.ir/) for their support and guidance regarding the need for open access initiatives in this
domain.

### Data Columns

Each Parquet file contains the following columns:

- **file name** (`string`): The unique identifier of the audio file.
- **transcript** (`string`): The ground-truth transcript corresponding to the audio.
- **duration** (`float64`): Duration of the audio file in seconds.
- **match quality** (`string`): Either "HIGH" for `CER < 0.05` or "MIDDLE" for `0.05 < CER < 0.2` between actual and hypothesis transcript.
- **hypothesis** (`string`): The best transcript generated by ASR as hypothesis to find the matching ground-truth transcript.
- **CER** (`float64`): The Character Error Rate (CER) of the ground-truth and hypothesis transcripts.
- **search type** (`int64`): Either 1 indicating the GT transcripts is result of Interval Search or 2 if a result of Gapped Search (refer to paper for more details).
- **ASRs** (`string`): The Automatic Speech Recognition (ASR) systems used in order to find a satisfying matching transcript.
- **audio** (`sequence`): The actual audio data.
- **samplerate** (`float64`): The sample rate of the audio.

## Usage

### Full Dataset
```python
from datasets import load_dataset
dataset = load_dataset("MahtaFetrat/Mana-TTS", split='train')
```

### Partial Download
To download only specific parts (e.g., for Colab/limited storage):
```bash
# Replace XX with part number (01, 02, etc.)
wget https://huggingface.co/datasets/MahtaFetrat/Mana-TTS/resolve/main/dataset/dataset_part_XX.parquet
```
**Streaming** (avoids full downloads):
```python
dataset = load_dataset("MahtaFetrat/Mana-TTS", streaming=True)
for sample in dataset["train"].take(100):  # Process batches incrementally
    print(sample)
```

## Citation

If you use Mana-TTS in your research or projects, please cite the following paper:

```bash
@inproceedings{qharabagh-etal-2025-manatts,
    title = "{M}ana{TTS} {P}ersian: a recipe for creating {TTS} datasets for lower resource languages",
    author = "Qharabagh, Mahta Fetrat  and Dehghanian, Zahra  and Rabiee, Hamid R.",
    booktitle = "Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)",
    month = apr,
    year = "2025",
    address = "Albuquerque, New Mexico",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.naacl-long.464/",
    pages = "9177--9206",
}
```

## License

This dataset is available under the cc0-1.0. However, the dataset should not be utilized for replicating or imitating the speaker's voice for malicious
purposes or unethical activities, including voice cloning for malicious intent.

## Collaboration and Community Impact
We encourage researchers, developers, and the broader community to utilize the resources provided in this project, particularly in the development of high-quality screen readers and other assistive technologies to support the Iranian blind community. By fostering open-source collaboration, we aim to drive innovation and improve accessibility for all.