Pazhvak is a publicly available, word-level Farsi speech corpus designed for research and development in Farsi speech processing. This dataset consists of 88,535 samples 📊, divided into training (80%) and validation (20%) splits. It includes 4,018 unique words 🔤 recorded by 61 speakers 🗣️ (38 male and 23 female). Each recording ranges from 0.5 to 7 seconds, sampled at 16 kHz in mono 🎧.

Quick Start 🚀

from datasets import load_dataset

ds = load_dataset("MASaraji/PAZHVAK")
print(ds["train"][0])

Dataset Structure 📁

Format: WAV audio files 🎵 + CSV Labels 🗂️
Sampling Rate: 16 kHz
Channel: Mono 🔈
Splits: Train (80%), Validation (20%)
Labels: Farsi Label 🇮🇷 + Finglish Label 🔡 + Phonetic Label 🔤

Processing Pipeline 🛠️

Our data preparation pipeline has three key stages: Pre-processing, Post-processing, and Enhanced Recording.

Pre-processing 🔍

Most frequent Farsi words were selected.
Words were evenly distributed among speakers 👥.

Post-processing 🧹

Corrupted or low-quality recordings were removed ❌.
All audio was resampled to 16 kHz 🔄.
Stereo channels were converted to mono 🔊.
Files were systematically renamed 🗃️.

Enhanced Recording 🎚️

Silence was trimmed ✂️, and a 0.5s silence was added to the beginning and end 🔇.
Bitrate was set to 192 kbps ⚙️.
Amplitude normalized to -20 dBFS 📶.

Evaluation with Whisper 🤖

To assess transcription quality, we evaluated Pazhvak using OpenAI's Whisper model:

Character Error Rate (CER): 35% 🧮

We further analyzed 50 words with the highest and lowest error rates 📉📈.

Words with multiple valid pronunciations or containing homophonic letters had the highest error rates ❗.

Common, unambiguous words saw significantly lower error rates ✅.

Words with multiple valid pronunciations or containing homophonic letters had the highest error rates.

Common, unambiguous words saw significantly lower error rates.

Diversity 🌍

Pazhvak prioritizes speaker diversity to ensure robustness across dialects, age groups, and genders:

61 speakers (38 male, 23 female) 🧔👩
Varied age ranges 📆
Regional and accentual diversity across Iran 📍

Each word in the corpus was segmented into individual phonemes, revealing a total of 36 distinct phonemes. While native Farsi phonology comprises 23 consonants and 6 vowels, the remaining phonemes originate from loanwords (e.g., Arabic and English), introducing non-native sounds into the corpus.

License 📜

This dataset is licensed under the MIT License ✅.

Citation 🧾

If you use Pazhvak in your research, please cite it as follows:

@article{PazhvakCorpus,
  title={PAZHVAK: A Word-Level Farsi Speech Corpus by University of Hormozgan},
  author={Mohammad Azim Saraji, Abdullah Khalili, Ahmad Hatam},
  year={2025}
}

Contact 📬

For questions, feedback, or collaborations, please reach out: 📧 [email protected]

Contributors

Arian Nazeri, Yasin Mohammadi, Amir Shakibafar, Alireza Keshavarz, Sepehr Simkhah , Parsa Gheibi, Mahshad Asadi, Maryam Eslami, Alireza Anoosheh, Vida Rezvani, Helia Attar, Ida Andishgar, Mahdi Sharifi, Mohammad Kazem Rahimi, Roghaye Molamohammadi, Alireza Dalir, Shakiba Pedram, Ashvagh Asnavi, Yasin Sanjari, Mobina Ehterami , Ava Ghiyasian, Arshia Alishahi, Mohammad Reza Razaghpoor, Bassam Nazemi, Fatemeh Safari, Sadra Kamyab, Mohammad Mohsenpoor, Paria Porandish, Fatemeh Ghorbanizadeh, Rasoul Tirandaz, Hamidreza Konarizadeh, Fatemeh Samadi, Fatemeh Dehmiani, Sana Ghiasi, Parisa Shekari, Fatemeh Raeisi, Alireza Ahmadiniya, Hasan Taghavi, Mehrshad Hajizadeh, Reza Rouhani, Mohammad Parvizi, Alireza Ashoori, Hoda Hoseini, Sayna Sayebani, Elaheh Azarakhshi, Elina Heydari, Yousef Bahrami, Erfan Ahrari, Sajjad Mohammadhoseinizadeh, Farzad Adelfar, Mohammad Amin Miri, Mohammad Motiei, Mohammad Moeini, Mohammadreza Ghasemi, Zahra Tanide, Mohammad Bagher Palahang, Asal Ameli, Tabasom Poriaei, MohammadAmin Johari, Hojjat Shahriari, Mohammad Ahadi, Ali Ashoorizadeh, Seyed Mohammad Mosavi, Mohammad Esmaeili, Kamand Kargar, Kimiya Mohazzabi, Neda Nazifi, Mojtaba Mehdipoor, Ali Touhidi

Downloads last month: 24