saeedzou/persianvox_all

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

PersianVox-All is the larger, less-filtered counterpart of PersianVox: a multi-speaker Persian (Farsi) speech corpus automatically mined from in-the-wild unlabeled data. It contains every utterance that passed language and speech-quality (MOS) filtering, without the additional dual-ASR transcript-agreement filtering applied to the main PersianVox release. It is therefore substantially larger, at the cost of some transcripts being less reliable.

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.

PersianVox-All was produced with the same fully automated pipeline used for PersianVox, which 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 flag unreliable transcriptions without any ground-truth transcripts.

Unlike the main PersianVox release, this dataset does not discard utterances based on that dual-ASR agreement check — it keeps every utterance that passed only the language-detection and MOS quality filters. The wer, cer, start_cer, and end_cer fields are still included so users can apply their own transcript-reliability threshold if needed.

Relationship to PersianVox

	PersianVox	PersianVox-All
Language filtering (lang == fa, lang_prob > 0.95)	✅	✅
MOS quality filtering (mos ≥ 3.5)	✅	✅
Dual-ASR agreement filtering (cer < 12.5, wer < 15, edge CER < 50)	✅	❌
Utterances	625,192	1,039,253

In short, PersianVox is a strict subset of PersianVox-All: every sample in PersianVox also appears in PersianVox-All, but PersianVox-All additionally includes samples whose two ASR hypotheses disagreed enough to fail the dual-agreement check. Use PersianVox for the highest-confidence transcripts, or PersianVox-All if you want maximum data volume and plan to apply your own filtering (e.g. via the included wer/cer/start_cer/end_cer fields).

Dataset Statistics

Statistic	Value
Utterances	1,039,253
Total duration	TODO: fill in total hours
Speakers	TODO: fill in speaker count
Vocabulary size	TODO: fill in vocab size
Mean utterance duration	TODO: fill in
Std. dev. of utterance duration	TODO: fill in

Supported Tasks

• Text-to-speech (TTS): training zero-shot / multi-speaker TTS models on audio paired with hyp_pc. Since dual-ASR filtering was not applied, consider filtering by wer/cer yourself for TTS training quality.
• Automatic speech recognition (ASR): the transcripts and quality fields make this corpus usable for ASR training/evaluation, including studying transcript disagreement itself.

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. Provided for reference/filtering — not used to filter this release.
cer	float64	Character Error Rate between hyp_pc and hyp_nopc. Provided for reference/filtering — not used to filter this release.
start_cer	float64	CER computed on just the first 7 characters of the utterance. Provided for reference/filtering — not used to filter this release.
end_cer	float64	CER computed on just the last 7 characters of the utterance. Provided for reference/filtering — not used to filter this release.
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

Only the following filters were applied to this release:

• lang == "fa" and lang_prob > 0.95 (language filtering)
• mos ≥ 3.5 (speech-quality filtering)

The dual-ASR agreement filtering used for the main PersianVox release (cer < 12.5, wer < 15, edge CER < 50) was not applied here — those fields are included in every row so you can apply the same or a custom threshold yourself.

Data Splits

Split	Examples
train	1,039,253

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 this corpus.

Considerations for Using the Data

• Because the dual-ASR agreement filter was not applied, this release contains a meaningfully higher proportion of unreliable or mismatched transcripts than PersianVox. Use the wer/cer/start_cer/end_cer fields to apply your own reliability threshold if transcript accuracy matters for your use case.
• Transcripts are machine-generated (ASR) rather than human-verified.
• 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}, 
}