Datasets:

Thomcles
/

Persian-Farsi-Speech

features:
  - name: audio
    dtype:
      audio:
        sampling_rate: 16000
  - name: sentence
    dtype: string
task_categories:
  - text-to-speech
  - automatic-speech-recognition
  - audio-to-audio
language:
  - fa
tags:
  - persian
  - farsi
  - speech
  - TTS
  - ASR
  - Iran
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
dataset_info:
  features:
    - name: sentence
      dtype: string
    - name: audio
      dtype:
        audio:
          sampling_rate: 16000
    - name: mos_p808
      dtype: float64
    - name: mos_sig
      dtype: float64
    - name: mos_bak
      dtype: float64
    - name: mos_ovr
      dtype: float64
  splits:
    - name: train
      num_bytes: 37113065696.67625
      num_examples: 109401
  download_size: 47491068931
  dataset_size: 37113065696.67625
license: cc-by-4.0
pretty_name: Persian Farsi Speech Dataset
size_categories:
  - 100K<n<1M

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.

📦 Data Sources

The following public datasets were used as base sources:

farsi_voice_dataset - Apache license 2.0
asr-farsi-youtube-chunked-10-seconds - No license
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
Audio quality labeling dnsmos

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.}
}