PersianVox: A Prosody-Aware Approach for Speech Dataset Generation from In-the-Wild Data

PersianVox is a 2,400-hour, multi-speaker Persian (Farsi) speech corpus automatically mined from in-the-wild unlabeled data. It is, to date, the largest open-source speech resource for Persian, built to support zero-shot text-to-speech (TTS) research and other speech tasks in low-resource-language settings.

Dataset Summary

Advancement of zero-shot text-to-speech synthesis is currently hindered for low-resource languages by the scarcity of large-scale, high-fidelity speech datasets. Traditional alignment-based methods require rare verbatim transcripts, while standard in-the-wild pipelines often rely on single-model automatic speech recognition and silence-based segmentation, leading to transcription errors and truncated prosody.

To address these challenges for the Persian language, PersianVox was produced with a fully automated pipeline that generates high-quality speech corpora from in-the-wild data. The pipeline integrates:

• A prosody-aware segmentation strategy that uses acoustic turn-detection to preserve linguistic completeness and optimize utterance duration for long-context modeling.
• A dual-model agreement mechanism, leveraging two distinct ASR model architectures to filter unreliable transcriptions without any ground-truth transcripts.

The result is a 2,400-hour multi-speaker dataset for Persian, along with the first comparative benchmark of speech-quality-assessment methods for the language (a human-annotated subset is released separately to support that benchmark).

Dataset Statistics

Statistic	Value
Utterances	625,192
Total duration	~2,408.67 hours
Speakers	3,248
Vocabulary size	212,216
Mean utterance duration	13.87 s
Std. dev. of utterance duration	6.32 s

Supported Tasks

• Text-to-speech (TTS): primary intended use — training zero-shot / multi-speaker TTS models on audio paired with hyp_pc.
• Automatic speech recognition (ASR): the transcripts and CER/WER quality fields make this corpus usable for ASR training/evaluation as well.

Dataset Structure

Data Instances

Each row is a single utterance: an audio clip, its transcript(s), quality/filtering scores, and provenance metadata linking it back to the source video and merged sub-segments.

Data Fields

Field	Type	Description
audio	Audio	The utterance's audio clip.
duration	float64	Duration of the clip, in seconds.
lang	string	Detected language; always "fa" — non-Persian samples were filtered out.
lang_prob	float64	Language-detection confidence from Whisper Large V3; all retained samples have lang_prob > 0.95.
mos	float64	Predicted Mean Opinion Score (speech quality), measured with SCOREQ. All samples with mos less than 3.5 were filtered out.
hyp_pc	string	Punctuated transcript hypothesis from the authors' proprietary ASR model. This is the transcript used for TTS training.
hyp_nopc	string	Transcript hypothesis from a second, punctuation-less ASR model, used as an independent check against hyp_pc.
wer	float64	Word Error Rate between hyp_pc and hyp_nopc, used for filtering.
cer	float64	Character Error Rate between hyp_pc and hyp_nopc, used for filtering.
start_cer	float64	CER computed on just the first 7 characters of the utterance.
end_cer	float64	CER computed on just the last 7 characters of the utterance.
channel_id	string	ID of the source channel the sample was drawn from.
video_id	string	ID of the source video.
speaker	string	Speaker identifier.
start	float64	Start time of the clip within the source video.
end	float64	End time of the clip within the source video.
eos_state	string	"complete" or "incomplete" — whether the utterance ends in a complete sentence, as judged by smart-turn-v3. A dedicated merging strategy built around this signal was used to increase the proportion of complete utterances.
segments	list	The sub-segments (each with start, end, speaker, index, eos_prob, eos_state) that were merged together to form this utterance.

Data Filtering

Samples were filtered based on empirically chosen thresholds (see paper for details):

• cer < 12.5
• wer < 15
• start_cer / end_cer (edge CER) < 50

Data Splits

Split	Examples
train	625,192

Dataset Creation

Source Data

Utterances were mined from unlabeled, publicly available web video/audio, identified by channel_id and video_id, and segmented using a prosody-aware, turn-detection-based strategy rather than simple silence-based segmentation.

Annotations

All transcripts and quality labels (hyp_pc, hyp_nopc, wer, cer, start_cer, end_cer, mos, eos_state, lang, lang_prob) were produced automatically by the pipeline described above — no human transcription was used to build the main corpus. A separate, human-annotated subset was additionally released to benchmark speech-quality-assessment methods for Persian (see paper).

Considerations for Using the Data

• Transcripts are machine-generated (ASR) rather than human-verified; use the provided wer/cer/mos fields to apply your own quality thresholds if the defaults don't suit your use case.
• Released under CC BY 4.0 — redistribution and commercial use are permitted with attribution.

License

This dataset is released under the CC BY 4.0 license.

Citation

If you use this dataset, please cite the accompanying paper:

@misc{zouashkiani2026persianvoxprosodyawareapproachspeech,
      title={PersianVox: A Prosody-Aware Approach for Speech Dataset Generation from In-the-Wild Data}, 
      author={Saeedreza Zouashkiani and Soheil Khalesi and Saman Soleimani Roudi and Sajjad Amini and Shahrokh Ghaemmaghami},
      year={2026},
      eprint={2609.19324},
      archivePrefix={arXiv},
      primaryClass={eess.AS},
      url={https://arxiv.org/abs/2609.19324}, 
}