Reza2kn/telegram-audiobook-chizzled

Telegram Persian Audiobook Chizzled

1,555,434 Persian audiobook clips · 14,430.384 hours · 16 kHz mono PCM WAV · public Parquet release

This is a large, provenance-preserving collection of Persian audiobook audio gathered from 26 Telegram channels accessible to the collector account. Each source message is retained as message-level provenance and segmented with Silero voice-activity detection (VAD) into pause-aware clips. The audio bytes are embedded in Parquet files, so the dataset can be streamed directly from the Hub.

Scale, stated carefully. At 14,430.384 released clip-hours, this is, to our knowledge, the largest publicly available Persian audio release by duration as of 2026-09-12. It is an audio-only, audiobook-derived collection—not a licensed, transcribed, speaker-consented corpus—and is therefore not directly comparable in intended use or legal status to curated ASR/TTS datasets. This claim is a duration comparison, not a claim of quality, annotation coverage, or unrestricted reuse.

Snapshot

Measure	Value
Published clips	1,555,434
VAD-cleaned clip duration	14,430.384 h (601.27 days)
Normal clips	1,206,482
Explicit short_tail clips	348,952
Telegram posts inspected	68,315
Audio messages processed	30,414
Decoded source-audio duration	15,719.309 h
No-speech messages	92 (19.542 h)
Malformed/failed source messages	10 (0.627 h)
Non-audio posts skipped	37,799
Audio storage	Embedded bytes in 6,349 Parquet shards

The duration_sec figure is the duration after VAD removal of non-speech intervals, not the original wall-clock duration. Consequently, clips bounded around the 30–45 second wall-clock policy can have shorter retained speech duration.

Relative scale

Public release	Released audio duration	This release relative to it
This dataset	14,430.384 h	1.00×
Shenava 1: visualears-persian-asr-16k	7,771.12 h	1.86×
ParsVoice	2,177.9 h	6.63×

The Shenava comparison uses the public-release duration reported in its project hub. The ParsVoice comparison uses its currently published Hugging Face card, which reports 1,360,521 segments and 2,177.9 hours. These are scale comparisons only: Shenava is an ASR-oriented release, while ParsVoice is a transcribed, quality-filtered TTS corpus.

Source channels

The collection was drawn from the following channels. Each row preserves channel, channel_id, message_id, and source_url, so individual source messages remain traceable.

Channel	Channel
1	@gondage
2	@asrebidaari
3	@podchi
4	@ketabkhaneh2024
5	@shokrgozarry
6	@ketabhaye_sotii
7	@naserzeraatiketabsoti
8	@PersianAudioLibrary
9	@Secret_group_audio_library
10	@ketttabhayesotii
11	@keetaab7
12	@ketabkhaneyesowti
13	@listen_to_my_audio_library
14	@audiobo0ok
15	@boooks1000
16	@mhrbaba
17	@ketabhaimarx
18	@Adathaeatomi_ekvanediv
19	@Jahanhayepichide
20	@Yervand_Abrahamian
21	@jozzazkol
22	@boook_audio
23	@dostoevsky_audio
24	@parvazamani
25	@books_markaz15
26	@ketabchi_co

Channel membership or public reachability does not establish ownership, copyright, narrator consent, or a redistribution license for an underlying recording.

Processing

1. The crawler inspected message history in the 26 channels and selected Telegram documents recognized as audio.
2. Input was decoded to mono 16 kHz audio.
3. Silero VAD used a speech threshold of 0.5; speech regions separated by up to 1.2 s were merged.
4. Adjacent VAD regions were grouped at pauses into chunks with a 30–45 s wall-clock target. Long uninterrupted speech runs were cut near a low-energy point.
5. The retained VAD regions were concatenated and written as mono 16 kHz, PCM-16 WAV. Long silence is removed; this is not sentence-level segmentation and does not guarantee a linguistic sentence boundary.
6. Audio and provenance were committed together in 250-row-or-smaller embedded-audio Parquet shards. A local WAV was deleted only after the corresponding Hub commit succeeded.

Production used VAD/chunking with denoising disabled. The historical chizzled field means the clip passed through this pipeline; it does not mean MP-SENet denoising was applied to this production release.

Schema

Each Parquet row includes embedded audio plus:

Field	Meaning
channel, channel_id, message_id, source_url	Telegram source identity and link
source_filename, source_sha256	Source-media filename and SHA-256
clip_key, clip_index, clip_sha256	Stable clip identity, order, and SHA-256
duration_sec	Retained, VAD-cleaned audio duration
wall_start_sec, wall_end_sec, intervals_json	Positions and retained VAD intervals in the decoded source
vad_threshold, vad_merge_gap_sec	VAD configuration
chizzled, status	Processing marker and done/short_tail status

Loading

from datasets import load_dataset

dataset = load_dataset(
    "Reza2kn/telegram-audiobook-chizzled",
    split="train",
    streaming=True,
)
sample = next(iter(dataset))
print(sample["channel"], sample["source_url"], sample["duration_sec"])

Streaming is recommended. The files are large and audio is embedded in the Parquet shards.

Intended use and limitations

• The release is useful for research on Persian speech representation learning, audiobook narration, VAD, segmentation, retrieval, and acoustic pretraining.
• It has no transcripts, no speaker labels, no sentence labels, and no quality labels. It is not directly TTS- or ASR-supervised training data.
• The audio is predominantly narrated audiobook material. It is not representative of conversational Persian, dialect coverage, or all demographics.
• VAD is a pause heuristic, not syntax or semantic understanding. A clip can still cross a sentence boundary or start/end near a word depending on recording acoustics.
• Some source messages were unreadable or contained no detected speech; counts are reported above rather than silently removed from the audit trail.
• Duplicate books, narrators, editions, or passages may remain across channels. No global deduplication, speaker diarization, transcription, or rights verification was performed.

Rights, attribution, and removal

This repository makes no claim of ownership over the underlying recordings, books, performances, or channel uploads. The source_url and SHA-256 fields are included to support provenance and removal review. The repository-level other license applies only to release metadata and does not grant rights to source recordings.

Use is restricted to purposes for which you have the necessary rights. If you are a rights holder or channel administrator and want a source removed, open a Hub discussion or contact the repository owner with the affected source_url or clip_key; affected material can be located and removed using the provenance fields.

Reproducibility

The crawl maintained a SQLite message/clip ledger, SHA-256 identifiers, a JSONL provenance manifest, and post-commit deletion accounting. The final audit verified that every current manifest key had a corresponding committed upload (manifest_not_uploaded = 0).