kiarashQ/farsi-asr-unified-cleaned

🎧 Farsi ASR Unified Dataset (Parquet Sharded Edition)

Overview

The Farsi ASR Unified Dataset is a large-scale, high-quality, and fully standardized collection of Persian (Farsi) speech-to-text data — designed specifically for modern machine learning and ASR (Automatic Speech Recognition) workflows.

This dataset consolidates audio–text pairs from multiple open sources, applies a rigorous cleaning and normalization pipeline, and stores everything efficiently in Parquet shards with embedded audio.

• Total Duration: ~1,392.8 hours
• Total Samples: 1,278,937
• Format: Parquet (sharded, with embedded 16 kHz audio)

---

✨ Highlights

This edition focuses on speed, simplicity, and reliability when working with large-scale speech data.

• 🚀 High Performance – Audio and transcripts are packed into ~500 MB Parquet shards, enabling fast reads and easy handling.
• 📦 Self-Contained – Each Parquet file includes both text and audio bytes. No need to manage separate audio folders.
• 💻 Streaming-Friendly – Optimized for the Hugging Face datasets library — load or stream directly without downloading the full dataset.
• ✅ Clean & Minimal – The repository only includes the data shards and this README, keeping it lightweight and easy to navigate.

Limitations and Known Issues

• Due to the nature of aggregating multiple large, web-scraped datasets, a very small fraction of samples may contain language mismatches. For example, some samples have been observed to contain Arabic audio with a Farsi translation as the transcript. The estimated rate of such errors is well below 0.01%, and is not expected to significantly impact the training of robust ASR models. Future versions of this dataset may include a language identification filtering step to address this.

---

🧠 Usage

You can load this dataset directly with the Hugging Face datasets library.

1️⃣ Installation

pip install datasets[audio]

Loading the Dataset

You can easily load the dataset (either streaming or fully downloaded) with a single line of code.

from datasets import load_dataset

# Stream the dataset (recommended for exploration)
ds = load_dataset("kiarashQ/farsi-asr-unified-cleaned", split="train", streaming=True)

# Or download the entire dataset
# ds = load_dataset("kiarashQ/farsi-asr-unified-cleaned", split="train")

# Example usage
example = next(iter(ds))
print(example)
# {'path': '...', 'text': '...', 'audio': b'...', 'sampling_rate': 16000}

# Decode the audio bytes
import datasets
ds = ds.cast_column("audio", datasets.Audio(sampling_rate=16000))

example_audio = next(iter(ds))
print(example_audio["audio"])
# {'path': None, 'array': array([...], dtype=float32), 'sampling_rate': 16000}

---

Dataset Structure

The repository contains a single data directory:

/shards/

Each file is a Parquet shard (~500 MB) containing a subset of the data:

fa_asr_unified_000001.parquet
fa_asr_unified_000002.parquet
...

Dataset Schema

Each row in the Parquet files contains the following fields:

Column Name	Type	Description
path	string	The original relative path of the audio file (e.g., cv22_fa/common_voice_fa_12345.wav). Useful for reference and traceability to the source data.
text	string	The final, cleaned and normalized transcript.
audio	binary	The raw bytes of the audio file (16kHz, mono, PCM WAV format). When using datasets.cast_column("audio", Audio()), this is automatically decoded into a NumPy array.
sampling_rate	int32	The sampling rate of the audio. This is consistently 16000 for all files in this dataset.

---

Processing & Filtering Pipeline

The underlying data was prepared using the following steps before being sharded into Parquet files:

1. Audio Standardization: All source audio was converted to a standard format: 16kHz, mono, PCM16 WAV.
2. Initial Deduplication: Entries were deduplicated based on the combination of (transcript text + audio content hash) to remove exact duplicates.
3. Advanced Text Normalization: Transcripts were cleaned and standardized:
   • Standardized punctuation (quotes, dashes, etc.).
   • Filtered to a character whitelist containing only the Persian alphabet, numbers, basic punctuation, and the Zero-Width Non-Joiner (\u200c). All other characters (e.g., English, emojis) were removed.
   • Normalized whitespace.
4. Initial Filtering:
   • Removed entries with empty transcripts after normalization.
   • Removed all audio clips with a duration < 1.0 second.
5. Outlier Filtering :
   • Max Duration: Kept samples with duration <= 30.0 seconds.
   • Character Rate: Kept samples with a character rate between 1.5 and 20.0 characters/second.

---

📚 Included Datasets

This unified dataset was created by aggregating and processing data from the following sources. Sincere thanks to all the original dataset authors for their contributions to the Farsi ASR ecosystem ❤️.

• CommonVoice_22_fa
• srezas/farsi_voice_dataset
• pourmand1376/asr-farsi-youtube-chunked-30-seconds
• MohammadGholizadeh/youtube-farsi
• MohammadGholizadeh/filimo-farsi
• MahtaFetrat/Persian-Tacotron2-on-ManaTTS
• saeedzou/PersianVox_NM
• vhdm/persian-voice-v1
• payamvha/northpersian_Dataset_TTS

---

🏁 Summary

This repository provides a clean, unified, and high-performance foundation for training and evaluating Farsi ASR models. Whether you’re fine-tuning Whisper, training from scratch, or benchmarking multilingual models, this dataset is ready to go — fast, consistent, and easy to use.

---

Maintained by kiarashQ

License: Refer to individual source dataset licenses.

---