ParsVoice

A Large-Scale Multi-Speaker Persian Speech Corpus for Text-to-Speech Synthesis

📣 Accepted to the EMNLP 2026 Main Conference.

ParsVoice is the largest publicly available Persian speech–text corpus tailored for training multi-speaker text-to-speech (TTS) systems. It is built from long-form Persian audiobook recordings using a fully automated pipeline combining sentence-aware segmentation, ASR transcription, a ParsBERT sentence-completion classifier, binary-search boundary optimisation, ECAPA-TDNN speaker identification, and Persian-specific audio and text quality assessment.

Dataset summary

Segments	1,360,521
Total duration	2,177.9 hours
Speakers (automatically inferred global IDs)	1,803
Source audiobooks	1,877
Audio format	FLAC, 16 kHz, mono, 16-bit (lossless)
Tokens / vocabulary	17.5 M / 224 K unique word forms
Mean segment length	5.79 s (median 4.58 s), 12.8 words
Language	Persian (Farsi)

Quick start

from datasets import load_dataset

# Streaming is strongly recommended — the full corpus is ~130 GB.
ds = load_dataset("MohammadJRanjbar/ParsVoice", split="train", streaming=True)

sample = next(iter(ds))
print(sample["transcript"])           # Persian transcript
print(sample["speaker_id"])           # global speaker ID
print(sample["audio"]["sampling_rate"])  # 16000
print(sample["audio"]["array"].shape)    # decoded waveform

Download the whole thing instead (needs ~130 GB of disk):

ds = load_dataset("MohammadJRanjbar/ParsVoice", split="train")

Dataset structure

The corpus ships as two configs. The default config holds what you need to train a TTS or ASR model; the metadata config holds every additional per-segment field, kept separate so the dataset viewer stays readable and so you can inspect quality statistics without downloading any audio.

default — audio, transcript, speaker

Field	Type	Description
segment_id	string	Stable unique segment identifier, {book_id}_{segment_number}
audio	Audio (16 kHz)	Speech segment, FLAC-encoded (lossless)
transcript	string	Persian transcript of the segment
speaker_id	string	Global speaker ID, consistent across the whole corpus

metadata — everything else (no audio, ~70 MB)

Load it on its own and join to the audio on segment_id:

meta = load_dataset("MohammadJRanjbar/ParsVoice", "metadata", split="train")
df = meta.to_pandas()          # ~1.36 M rows, downloads in seconds

# e.g. keep only complete sentences from high-quality, confidently-clustered audio
sel = df[(df.completion_status == "Complete") &
         (df.audio_quality_score >= 90) &
         (df.speaker_confidence >= 0.6)]
print(len(sel), "segments,", sel.duration_sec.sum() / 3600, "hours")

Field	Type	Description
segment_id	string	Join key against the default config
book_id	string	Hashed source-audiobook identifier (see Anonymisation)
narrator_id	string	Hashed narrator identifier; null when the source metadata had no narrator name
narrator_known	bool	Whether this segment's narrator is named in the source metadata
speaker_id	string	Global speaker ID (same as in default)
local_speaker_id	int32	Within-book speaker cluster index
speaker_confidence	float32	Clustering confidence for the speaker assignment
duration_sec	float32	Segment duration in seconds
sampling_rate	int32	Always 16000
book_offset_sec	float64	Start time of this segment within the original audiobook
audio_quality_score	float32	Composite audio quality, 0–100
snr_db	float32	Estimated signal-to-noise ratio
dynamic_range_db	float32	Dynamic range
clipping_percentage	float32	Percentage of clipped samples
silence_percentage	float32	Percentage of silence
spectral_centroid_mean	float32	Mean spectral centroid
spectral_rolloff_mean	float32	Mean spectral rolloff
zero_crossing_rate	float32	Zero-crossing rate
mfcc_variance	float32	MFCC variance
has_background_music	bool	Background music detected (inaSpeechSegmenter)
is_clean	bool	Passed the combined cleanliness check
completion_status	string	Complete / Incomplete / Invalid, from the ParsBERT sentence-completion classifier
number_of_extensions	int32	Boundary-extension iterations applied during segmentation
start_trimmed_ms	int32	Audio trimmed from the segment start by boundary optimisation
end_trimmed_ms	int32	Audio trimmed from the segment end
original_duration_ms	int32	Duration before boundary optimisation
final_duration_ms	int32	Duration after boundary optimisation
trim_method	string	Boundary-optimisation method used

Filtering guidance

The release is deliberately inclusive so you can apply your own thresholds rather than inherit ours. Useful cuts, computed over the released rows:

Filter	Segments	Share
All released segments	1,360,521	100%
completion_status == "Complete"	1,250,325	91.9%
audio_quality_score >= 75	1,352,643	99.4%
audio_quality_score >= 90	1,229,811	90.4%
snr_db >= 20	1,352,614	99.4%
has_background_music == False	1,334,947	98.1%
narrator_known == True	1,131,531	83.2%

For TTS training we recommend at minimum completion_status == "Complete" and audio_quality_score >= 75. A small number of segments are unusually long (49,301 exceed 15 s, 837 exceed 60 s) or very short (1,447 below 1 s); filter on duration_sec to suit your model's batching. audio_quality_score is null for 6,431 segments whose quality metrics were not recorded upstream.

Segments with narrator_known == True are the ones whose speaker identity is corroborated by narrator metadata — restrict to these if you need speaker labels that are not purely the product of automatic clustering.

Anonymisation

Book titles and narrator names are not distributed. Each is replaced by a salted SHA-256 identifier (book_id, narrator_id), truncated to 16 hex characters. The salt is held privately by the authors and is not published, so the hashes cannot be reversed or brute-forced against a catalogue of Persian audiobook titles. The identifiers remain stable and consistent across the corpus, so you can still group segments by book or by narrator, and count distinct books and narrators, without recovering who or what they are.

Relationship to the paper

The paper reports a TTS-ready subset of 1,364,671 segments totalling 2,199.7 hours. This release contains 1,360,521 segments (2,177.9 hours): 4,150 segments were dropped because their per-segment metadata could not be recovered, so they could not be described or filtered here. All other figures (17.5 M tokens, 12.8 mean words per segment, ~1,800 speakers) match the paper closely.

Note that, unlike the filtered subset described in the paper, this release retains segments flagged Incomplete (110,182) and the small number below the audio-quality threshold, so that users can choose their own thresholds. Apply the filters above to reproduce the paper's TTS-ready configuration.

Corpus construction

Full details are in the EMNLP 2026 paper; the pipeline is on GitHub. In brief:

1. Segmentation — WebRTC VAD (aggressiveness 0) proposes silence-based boundaries.
2. Transcription — each candidate segment is transcribed with a Persian ASR backend.
3. Completeness validation — a ParsBERT classifier (97.4% F1) flags incomplete sentences; those segments are iteratively extended in 0.1 s steps and re-transcribed.
4. Boundary optimisation — binary search finds the largest trim at each boundary that leaves the transcription character-for-character identical, removing leading silence and trailing artifacts.
5. Quality assessment — composite audio scoring (SNR, dynamic range, clipping, silence, background music) and a Persian-specific text quality framework.
6. Speaker identification — ECAPA-TDNN embeddings clustered within each book, then merged across the corpus into global speaker IDs.

Transcript accuracy was independently audited: against human reference transcriptions of 500 randomly sampled segments, ParsVoice transcripts achieve 4.90% WER and 1.81% CER, with 69.0% of segments transcribed perfectly.

Validation

Fine-tuning XTTSv2 on ParsVoice — operating directly on raw Persian text with no phoneme front-end — yields, on unseen reference speakers:

Metric	Score
Naturalness (MOS)	3.60 ± 0.09
Speaker similarity (SMOS)	4.03 ± 0.08
Intelligibility (MOS)	4.03 ± 0.08
Speaker similarity (ECAPA-TDNN cosine)	80.0%

Limitations

• Domain. Audiobooks only, so the speaking style is formal and narrative. Spontaneous conversational speech is not represented.
• Transcripts are ASR-derived. No reference book texts were used; a small residual error rate remains despite multi-stage filtering.
• Speaker labels are automatic. Global speaker IDs come from ECAPA-TDNN clustering, not manual annotation. Treat them as automatically inferred identities; use narrator_known and speaker_confidence to restrict to better-supported labels.
• Gender imbalance. Among audiobooks with narrator metadata, roughly 33% of narrators are female and 67% male; ~40% of audiobooks lack narrator metadata entirely, so the full-corpus distribution is only partially observed.

Licensing and intended use

Annotations (transcripts, speaker IDs, quality scores) are released under CC BY-NC 4.0. The pipeline code is Apache 2.0. Audio segments are distributed for non-commercial research only, under gated access.

ParsVoice is derived from publicly accessible audiobooks on IranSeda. Complete recordings are not redistributed — only short, quality-filtered excerpts — consistent with Article 7 of Iran's Copyright Act, which permits quotation from published works for scientific and educational purposes with attribution. Copyright in the underlying recordings remains with IranSeda and the narrators; no ownership is claimed. Rights holders may request removal, and affected audio will be excluded from future releases.

Because ParsVoice can support voice-cloning-capable TTS, potential misuse includes impersonation, fraud, deceptive synthetic media, and disinformation. These are outside the intended scope of the dataset. Downstream users should obtain appropriate speaker consent for any deployed cloned voice and apply provenance or watermarking to publicly released synthetic speech.

Citation

@inproceedings{ranjbar2026parsvoice,
  title     = {ParsVoice: A Large-Scale Multi-Speaker Persian Speech Corpus for Text-to-Speech Synthesis},
  author    = {Ranjbar Kalahroodi, Mohammad Javad and Faili, Heshaam and Shakery, Azadeh},
  booktitle = {Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
  year      = {2026},
  note      = {Main Conference},
  url       = {https://arxiv.org/abs/2510.10774}
}

Authors

Mohammad Javad Ranjbar Kalahroodi, Heshaam Faili, Azadeh Shakery — School of Electrical and Computer Engineering, University of Tehran; Institute for Research in Fundamental Sciences (IPM).