Dataset Card for PSRB (Persian Speech Recognition Benchmark) - 1-Hour Sample

Dataset Summary

The Persian Speech Recognition Benchmark (PSRB) is a comprehensive dataset designed to evaluate Persian Automatic Speech Recognition (ASR) systems under diverse real-world conditions.
This 1-hour sample provides a representative subset of the full PSRB corpus, capturing various accents, speech styles, speaker demographics, and acoustic environments.

Supported Tasks and Leaderboards

• Automatic Speech Recognition (ASR)

Languages

• Persian (Farsi)

Dataset Structure

Data Instances

Each data instance in this sample is structured as:

{
  "audio_path": "file1.wav",
  "text": "می‌گم نمی‌خواین طبق نقشه جلو بریم؟ نقشه رو بی‌خیال غمت نباشه ما مستر رد پا رو داریم. حس بویایی و شم‌ام اشتباه نمی‌کنه. همین الان اشتباه کرده.",
  "audio_duration": 11.88,
  "number_of_speakers": 3,
  "gender": "male",
  "age": "mix",
  "accents": "standard",
  "formality": "informal",
  "semantic_content": "artistic&literary",
  "data_source": "animation",
  "acoustic_environment": "noisy",
  "spontaneous": 1
}

Data Fields

• audio_path: Path to the .wav audio file.
• text: Transcription of the audio in Persian.
• audio_duration: Duration of the audio file in seconds.
• number_of_speakers: Number of speakers present in the clip.
• gender: Gender of the speaker(s).
• age: Age category (e.g., child, teen, adult, senior, or mix).
• accents: Regional accent of the speaker or "standard."
• formality: Formality level ("formal" or "informal").
• semantic_content: Semantic topic or domain of the speech (e.g., artistic&literary, technological, medical).
• data_source: Source type of the data (e.g., animation, podcast, lecture).
• acoustic_environment: Recording environment (e.g., clean, noisy).
• spontaneous: 1 if speech is spontaneous, 0 if scripted.

---

Dataset Creation

Curation Rationale

The PSRB dataset was created to address the lack of comprehensive Persian ASR resources, covering linguistic diversity (accents, formality) and acoustic variability (clean, noisy, phone calls).

Source Data

Data sources include:

• News broadcasts
• Movies and TV shows
• Podcasts
• Lectures
• Audiobooks
• Talk shows

Collected from platforms such as Telewebion, Aparat, YouTube, and Iranseda.

Annotations

• Transcriptions manually created by expert native Persian speakers.
• Strict two-pass quality control review to ensure consistency and correctness.
• Rich metadata labeling for speaker demographics and speech conditions.

Personal and Sensitive Information

• All data was anonymized to protect the identity of participants.
• No personally identifiable information (PII) is present in this dataset.

---

Considerations for Using the Data

Limitations

• This sample may not capture the full variability of the complete PSRB corpus.

---

Additional Information

Licensing Information

This dataset is made available for research and educational purposes only.

---

Citation Information

If you use this dataset in your research, please cite:

@misc{psrb2025,
  title={PSRB: A Comprehensive Benchmark for Evaluating Persian Automatic Speech Recognition Systems},
  author={Nima Sedghiye and Sara Sadeghi and Reza Khodadadi and Farzin Kashani and Omid Aghdaei and Somayeh Rahimi and Mohammad Sadegh Safari},
  year={2025},
  publisher={Part AI Research Center},
  note={Preprint}
}