Persian-Farsi-Speech

# Persian (Farsi) TTS Dataset ## 🗂️ Dataset Description This dataset is a **Persian (Farsi) text-to-speech (TTS)** corpus built by **concatenating and denoising multiple existing Farsi datasets**. It is intended for training and evaluation of **speech synthesis (TTS)** models in Persian. Since the basic datasets were contaminated with unintelligible audio, I used **dnsmos** to keep only clean audio (mos_ovr >= 3.0, same value as for the Emilia dataset). The dataset contains two main columns: | Column | Description | |---------|--------------| | `audio` | Path or reference to the audio file (typically `.wav`) | | `sentence` | The corresponding Farsi text transcription of the audio | | `p808_mos` | Mean Opinion Score (MOS) predicted using the ITU P.808 standard for subjective speech quality evaluation. | | `mos_sig` | MOS estimating the perceived quality of the speech signal itself (clarity and naturalness of the voice). | | `mos_bak` | MOS estimating the quality of the background (amount of background noise and how intrusive it is). | | `mos_ovr` | Overall MOS combining both speech and background quality into a single overall quality score. | All data have been **normalized, denoised, and quality-checked** to ensure consistency across sources. --- <div align="center"><img width="400px" src="https://www.shutterstock.com/image-vector/persian-typography-iranian-art-translated-600nw-2053950194.jpg" alt="Iranian art" /></div> ## 📦 Data Sources The following public datasets were used as base sources: - [farsi_voice_dataset](https://huggingface.co/datasets/srezas/farsi_voice_dataset) - Apache license 2.0 - [asr-farsi-youtube-chunked-10-seconds](https://huggingface.co/datasets/pourmand1376/asr-farsi-youtube-chunked-10-seconds) - No license - [asr-farsi-youtube-chunked-30-seconds](https://huggingface.co/datasets/pourmand1376/asr-farsi-youtube-chunked-30-seconds) - Apache license 2.0 --- ## 🧹 Denoising & Preprocessing Audio preprocessing and denoising were performed using the following model: - **Denoising Model:** [sepformer](https://huggingface.co/speechbrain/sepformer-whamr16k) - **Audio quality labeling** [dnsmos](https://lightning.ai/docs/torchmetrics/stable/audio/deep_noise_suppression_mean_opinion_score.html) Other preprocessing steps may include: - Audio resampling to 16 kHz - Loudness normalization - Filtering out bad audio --- ## 📊 Dataset Statistics (to fill in) Basic dataset : | Field | Value | |--------|--------| | Number of samples | 236255 | | Total duration | 695 hours | | Average duration | 10.6 seconds per audio | | Sampling rate | 16 000 hz| | Compute cost | $75 (This may vary depending on the hardware.) | Filtered dataset: | Field | Value | |--------|--------| | Number of samples | 109401 | | Total duration | 417 hours | | Average duration | 13.7 seconds per audio | | Sampling rate | 16 000 hz| | Compute cost | $75 (This may vary depending on the hardware.) | --- ## ⚖️ License This dataset is licensed under CC BY 4.0. You can use it freely, including for commercial purposes. However, you must cite the dataset and its author. --- ## 🙏 Acknowledgments Thanks to the creators of the original Farsi speech datasets that made this compilation possible. --- ## 📬 Citation If you use this dataset, please cite it as: ``` @dataset{thomcles_2025_persian_farsi_dataset, title = {Persian-Farsi-Dataset}, author = {Thomcles}, year = {2025}, url = {https://huggingface.co/datasets/Thomcles/Persian-Farsi-Speech}, note = {A cleaned and concatenated dataset for Persian text-to-speech tasks.} } ``` ---