PersianVox_HB

PersianVox_HB is a high-quality, multispeaker Persian speech dataset derived from audio recordings of the Holy Bible. The dataset is sourced from bible.com (PCB=49.85 hours, TPV=71.4 hours), bible.is (NMV=24.17 hours), and wordproject.org (PHB=81.92 hours).

📚 Dataset Summary

Language: Persian (Farsi)
Speakers: Multiple speakers
Total Duration: 227.34 hours
Recording Sources:
- bible.com (PCB: 49.85 hours, TPV: 71.4 hours)
- bible.is (NMV: 24.17 hours)
- wordproject.org (PHB: 81.92 hours)
Domain: Religious texts (Holy Bible)
Alignment Checked With:
- In-house Whisper model fine-tuned from whisper-large-v3
- In-house NVIDIA FastConformer fine-tuned from NVIDIA FastConformer-Hybrid Large (fa)
- Public wav2vec2_v3

🔑 Key Descriptions

The dataset entries are stored as dictionaries with the following keys:

🧩 Common Keys

Key	Description
`audio_filepath`	Path to the audio file (usually relative to the dataset root)
`duration`	Length of the audio in seconds
`text`	Normalized transcript with normalization, space correction that is used to calculate metrics
`text_no_preprocessing`	Transcript before any text cleaning, normalization, or punctuation adjustments
`text_normalized`	Normalized version of the transcript, suitable for ASR training or evaluation
`score`	Average confidence score indicating alignment quality in the log space. Utterances with scores lower than -2 are dropped
`start_abs`	Average absolute amplitude of the beginning portion of the audio segment (used to evaluate energy or silence at the start)
`end_abs`	Average absolute amplitude of the ending portion of the audio segment (used to evaluate trailing silence or noise)
`tokens`	Number of tokens in text (BPE)

🤖 Model-Specific Keys

For each ASR model used to evaluate the audio, the following metrics are added, prefixed by the model name.
For example, for the NVIDIA FastConformer model, you'd have keys like cer_nvidia_fastconformer, wer_nvidia_fastconformer, etc.

Key Template	Description
`pred_text_{model}`	Raw predicted transcript from the ASR model
`pred_text_normalized_{model}`	Normalized-space-corrected version of the predicted transcript used to calculate metrics
`cer_{model}`	Character Error Rate between prediction and ground truth
`wer_{model}`	Word Error Rate between prediction and ground truth
`ins_rate_{model}`	Insertion rate — proportion of extra characters inserted
`del_rate_{model}`	Deletion rate — proportion of characters the model missed
`sub_rate_{model}`	Substitution rate — proportion of characters wrongly predicted

🔄 Licensing

This dataset is released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) License, as the source materials do not permit commercial use. You must credit the original sources (bible.com, bible.is, wordproject.org) when using or sharing this dataset.

🧠 Intended Use and Limitations

This dataset is ideal for:

Training and benchmarking ASR systems in Persian
Research on model robustness via ASR scoring
Speech alignment evaluation for religious texts

🚫 Ethical Notice

Please do not use this dataset to clone or imitate the speakers’ voices for malicious or unethical purposes. Respect the non-commercial restrictions of the source materials.

✉️ Citation

To cite this dataset:

@misc{persianvox_hb,
  title = {PersianVox_HB: A multispeaker Persian speech dataset from the Holy Bible},
  author = {Saeedreza Zouashkiani},
  year = 2025,
  url = {https://huggingface.co/datasets/persianvox_hb},
  note = {Derived from bible.com, bible.is, and wordproject.org, licensed under CC BY-NC 4.0}
}

Datasets:

saeedzou
/

PersianVox_HB

You need to agree to share your contact information to access this dataset