risaleinur/risale-nur-audio

Risale-i Nur Audio–Text Corpus

Gerçek insan okumalarını, aynı satırdaki kaynak metinle birlikte sunan açık bir ses–metin veri kümesidir. Yeni varsayılan audio-text yapılandırması 15 kitaptan 91.792 oynatılabilir klip ve 203,02 saat ses içerir. Metinler kanonik kaynaktan değiştirilmeden alınır ve her kayıt byte-exact section_id alıntılarıyla bağlanır.

An open speech corpus pairing human readings with their source text in the same row. The default audio-text config contains 91,792 playable clips (203.02 hours) from 15 books. Text is preserved from the canonical source and linked through byte-exact section_id citations.

Yapılandırmalar · Configs

Config	İçerik · Contents
audio-text (default)	91.792 embedded 16 kHz clips; canonical text; train/validation/test
tts-premium	50.804 quality-gated 24 kHz single-reader clips (~121 h)
clips	2.429-recording source catalogue
segments	107.403 source alignment rows + acoustic verification status / verified times
page-text	1.422 clip-to-source-text rows
page-index	5.938 source-text entries

audio-text

This config is intended for ASR, speech–text retrieval, read-along systems and non-Qur'anic TTS research. It includes three readers and:

Dil · Language	Rows
Turkish (tr)	87,475
Arabic (ar)	4,247
Persian (fa)	70

Splits are connected-component based. A source recording, canonical text hash or section_id cannot occur in more than one split.

Split	Rows	Hours
train	80,856	178.34
validation	4,305	9.62
test	6,631	15.06

Main fields:

• audio: embedded PCM WAV, mono 16 kHz
• text: canonical text corresponding to the reading
• book_id, book_name, reader, language, duration
• record_id, source_audio_id
• citations: byte-exact {section_id, quote} evidence
• match_method: source_exact or unicode_canonical_equivalent
• audio_sha256, text_sha256: integrity hashes

The 15 audio-side source texts and the canonical text corpus are byte-identical. All 91,792 published rows were independently re-derived: citation mismatch 0, audio-hash mismatch 0, page metadata keys 0, and cross-split source/text/section overlap 0. Another 3,113 cuts were excluded because their canonical location was missing, non-unique or not exact enough. Every clip is acoustically verified with Whisper against its text. 45,797 clips use the source timestamps; 13,214 clips whose source timestamps were shifted were re-aligned with Whisper word timestamps, re-cut and verified again; 32,781 clips come from source rows that had no timestamps at all (unmatched) and were aligned the same way, then linked to the canonical text through the same byte-exact citation gate. Excluded: 2,820 acoustically unverifiable, 1,888 without a unique canonical match, 213 that would link two splits. Alignment labels are machine-generated; they are not a substitute for a fully human-transcribed speech benchmark.

15 ses-kaynağı metni ile kanonik metinler byte düzeyinde aynıdır. Yayımlanan 91.792 satırın tamamı bağımsız yeniden türetildi: alıntı ve ses hash hatası 0, sayfa metaverisi 0, split'ler arası kaynak/metin/bölüm kesişimi 0. Yeri kesin kanıtlanamayan 3.113 kesit yayına alınmadı. Her klip Whisper ile dinlenip metniyle doğrulanmıştır: 45.797 klip kaynak zamanlarıyla, kaynak zamanı kaymış 13.214 klip Whisper kelime zamanlarıyla yeniden hizalanıp kesilerek; kaynakta hiç zamanı olmayan (unmatched) satırlardan 32.781 klip aynı yöntemle hizalanıp aynı byte-exact alıntı kapısından geçirildi. Çıkarılan: akustik doğrulanamayan 2.820, kanonik eşleşmesi tekil olmayan 1.888, split'leri birbirine bağlayacak 213. Hizalama etiketleri makine üretimidir; tamamen insan transkripsiyonlu bir benchmark olarak değerlendirilmemelidir.

Kullanım · Usage

from datasets import Audio, load_dataset

ds = load_dataset(
    "risaleinur/risale-nur-audio",
    "audio-text",
    split="train",
).cast_column("audio", Audio(decode=False))

row = ds[0]
print(row["text"], row["citations"])
audio_bytes = row["audio"]["bytes"]

tts-premium remains available for the earlier quality-gated, single-reader TTS view. clips, segments, page-text and page-index are metadata-oriented views retained for compatibility and corpus exploration. segments keeps the raw source alignment (start/end, some shifted) and adds verification (original_verified, realigned_verified, whisper_aligned, rejected, unverified) with verified_start/verified_end; use the verified times for cutting audio.

Kaynak, lisans ve sınırlar · Source, license and limits

Kaynak kayıt ve metinler RNK Neşriyat'tan yazılı izinle alınmıştır. Veri kümesi atıf şartıyla ticari olmayan kullanıma açıktır; ticari kullanım ayrı yazılı izin gerektirir. Okuyucu sesini rızasız klonlama/taklit etme ve yapay Kur'an tilaveti üretme lisans kapsamında değildir. Ayrıntılar: LICENSE ve NOTICE.md.

Source recordings and text are from RNK Neşriyat and are included with written permission. The dataset is available for non-commercial use with attribution; commercial use requires separate written permission. Unauthorized voice cloning or imitation and synthetic Qur'an recitation are outside the license. See LICENSE and NOTICE.md.

The aligned core is dominated by one reader; Arabic and Persian are much smaller than Turkish. Report per-language and per-reader results when evaluating general ASR performance.

Sürüm · Version

• v1.6.0 (2026-09-26): source rows that had no timestamps (unmatched) were aligned with Whisper word timestamps between verified anchors, re-cut, acoustically verified and matched to the canonical text: audio-text 58,858 → 91,792 (+32,934), tts-premium 41,089 → 50,804 (+9,715). New rows are in *-added-*.parquet shards (same schema). segments.verification gains whisper_aligned with verified times.
• v1.5.0 (2026-09-25): recovery. Clips removed in v1.4.0 were re-aligned with Whisper word timestamps between verified anchors, re-cut and re-verified: audio-text 45,797 → 58,858 (+13,061), tts-premium 32,158 → 41,089 (+8,931). Recovered rows are in *-recovered-*.parquet shards with the same schema and original split assignment. segments gains verification columns.
• v1.4.0 (2026-09-25): acoustic verification fix. Some source alignment files had shifted timestamps, so the audio of those clips read a different sentence. Every clip was re-checked with Whisper against its text; mismatched clips were removed: audio-text 61,831 → 45,797, tts-premium 43,956 → 32,158. Remaining rows, their audio, text and split assignment are unchanged.
• v1.3.0 (2026-07-29): added the canonical, page-free audio-text config with 61,831 embedded audio–text pairs and byte-exact section citations.
• v1.2.0 (2026-06-30): public non-commercial release.

Citation metadata is available in CITATION.cff. Corrections can be reported through the dataset Community tab.