risaleinur/risale-nur-audio

🎙️ Risale-i Nur Sesli Külliyat

Risale-i Nur Audio Corpus — spoken readings + Qur'an & Hizbü'l-Hakaik recitation

An open Turkish / Arabic speech corpus for ASR · TTS · forced-alignment Part of the Şahsı Manevi dataset family

Source · Kaynak: RNK Neşriyat

TL;DR — An open, freely downloadable speech corpus of spoken Risale-i Nur (Bediüzzaman Said Nursî), plus Qur'an and Hizbü'l-Hakaik (evrad) recitation. Five configs: a playable tts-premium set (43,956 quality-gated single-voice clips, ~103 h, embedded 24 kHz audio, split train/validation/test); clips (catalogue) and segments (107,403 forced-aligned spans; 66,616 matched); plus page-text (1,422 clip→source page text, incl. 631 unaligned readers) and page-index (5,938 pages). The single-voice core is a rare, large Turkish TTS asset.

TL;DR (TR) — Risale-i Nur'un sesli okumaları + Kur'an ve Hizbü'l-Hakaik tilavetlerinden açık (public), serbestçe indirilebilir bir korpus. Beş config: oynatılabilir tts-premium (43.956 kalite-süzülmüş tek-ses klip, ~103 sa, gömülü 24 kHz ses, train/validation/test); clips (katalog) ve segments (107.403 hizalı parça; 66.616 matched); ayrıca page-text (1.422 klip→sayfa metni, 631 hizalanmamış dahil) ve page-index (5.938 sayfa). Tek-ses çekirdeği nadir ve büyük bir Türkçe TTS varlığı.

📌 At a glance

▶️ Playable in this repo	tts-premium (default) — 43,956 clips · ~103 h · embedded 24 kHz mono audio · single voice (A.Köseoğlu) · Turkish · quality-gated (tier S/A)
📋 Metadata in this repo	clips (2,429 catalogue · training_grade) · segments (107,403 aligned spans; 66,616 matched) · page-text (1,422 clip→source page text, incl. 631 unaligned · leakage_group) · page-index (5,938 pages)
🗂️ Source corpus (context)	2,429 recordings · 462.1 h · 21 books · 26 reciters · 791 alignment files (raw per-clip mp3s not bundled — see clips.audio_url)
🌍 Languages	Turkish (primary) · Arabic (Qur'an / Hizbü'l-Hakaik / embedded âyât) · Persian · Kurdish (2 clips)
🎯 Tasks	ASR · TTS · forced-alignment · audio-classification
🔓 Access	Public · free download · license: other — RNK attribution, non-commercial · maintainer Şahsı Manevi
🏷️ Version	v1.2 · 2026-06

What ships here: ready-to-use audio is the tts-premium parquet (gold single-voice clips). clips/segments are metadata (full catalogue + all alignment spans); the raw multi-reader mp3 corpus is referenced by audio_url, not bundled. · Bu repoda hazır ses = tts-premium; clips/segments metadata; ham çok-okuyucu mp3 korpusu paketlenmedi (kaynağa audio_url ile işaret eder).

---

1 · What is this?

The Risale-i Nur Audio Corpus is an open Turkish/Arabic speech dataset of spoken readings of Bediüzzaman Said Nursî's Risale-i Nur Külliyatı, together with Qur'an recitation and Hizbü'l-Hakaik (evrad) recitation, by 26 human reciters. It exists to support ASR, TTS, forced-alignment research and to widen audio accessibility (e.g. read-aloud apps), as the audio member of the Şahsı Manevi dataset family.

🇹🇷 Türkçe

Risale-i Nur Sesli Külliyat, Bediüzzaman Said Nursî'nin Risale-i Nur Külliyatı eserlerinin sesli okumaları ile Kur'an ve Hizbü'l-Hakaik (evrad) tilavetlerinden oluşan, 26 okuyucunun seslendirdiği, Türkçe/Arapça açık (public) bir konuşma korpusudur. Amaç: ASR, TTS, hizalama ve erişilebilirlik. Şahsı Manevi veri ailesinin ses üyesidir.

2 · Supported tasks

• Text-to-Speech (TTS) — non-Qur'anic content. The playable tts-premium set (~103 h, single voice A.Köseoğlu, 24 kHz) is a rare, large, clean single-speaker Turkish corpus for TTS research, pronunciation modeling and accessibility (≈4.3× LJSpeech's 24 h after cleaning). Cloning or reproducing the reciter's identifiable voice is out of scope (needs his own consent — see §8).
• Automatic Speech Recognition (ASR) — Turkish with embedded Qur'anic Arabic and occasional Persian; strong material for Turkish↔Arabic code-switching. Use segments (text+timestamps) as labels; for multi-reader audio fetch source mp3s via clips.audio_url.
• Forced alignment / linguistics — 107,403 page-tagged spans with matched/unmatched confidence.
• Audio classification — quality tiering (qtier/snr_db on tts-premium); language/reciter analysis from clips/segments metadata (+ source audio).

🇹🇷 Türkçe

• TTS — Kur'an dışı. Oynatılabilir tts-premium (~103 sa, tek ses A.Köseoğlu, 24 kHz) — TTS araştırması, telaffuz ve erişilebilirlik için temiz tek-konuşmacı korpusu. Okuyucu sesini klonlama/yeniden üretme kapsam dışı (rıza gerekir — bkz. §8).
• ASR — Türkçe + gömülü Kur'an Arapçası ve nadir Farsça; tr↔ar code-switching için güçlü. Etiket olarak segments; çok-okuyucu ses için clips.audio_url.
• Hizalama / dilbilim — 107.403 sayfa-etiketli parça.
• Ses sınıflandırma — kalite kademesi (qtier/snr_db); dil/okuyucu analizi metadata + kaynak ses ile.

3 · Languages

segments language distribution (ISO-639-3):

Language	All segments	Matched (training-grade)
Turkish (tur)	98,281	61,763
Arabic (ara) — embedded âyât/evrâd	8,945	4,783
Persian (fas)	177	70
Total	107,403	66,616

• tts-premium audio is Turkish single-voice only (A.Köseoğlu).
• Kurdish (ku) appears at clip level only — 2 recordings in the Diğer Lisanlar book (Risaleya Nexweşan / Hastalar Risalesi); no aligned segments. It is reflected in the YAML language list (tr/ar/fa/ku).

Segment dili ISO-639-3 (tur/ara/fas); YAML language 2-harfli (tr/ar/fa/ku). tts-premium Türkçe tek-ses; Kürtçe yalnız 2 klip (digerlisan).

4 · Dataset structure

risale-nur-audio/
├── tts-premium/*.parquet     # ▶️ PLAYABLE — gold single-voice clips, embedded 24 kHz Audio  (default config)
├── clips/clips.parquet         # 📋 catalogue (2,429 recordings) + training_grade flag  (metadata)
├── segments/segments.parquet   # 📋 107,403 aligned spans: text/page/language/timestamps  (metadata)
├── page-text/clip_page_text.parquet  # 📋 1,422 clips → source page text (incl. 631 unaligned) + leakage_group
├── page-index/page_index.parquet     # 📋 5,938 pages → source text + n_clips_covering
├── DATASHEET.md              # Datasheets-for-Datasets (composition, collection, uses, rights ledger)
├── NOTICE.md  LICENSE  CITATION.cff  CONTRIBUTING.md  CODE_OF_CONDUCT.md
├── manifests/                # stats.json (single source of truth) + CHECKSUMS.sha256
└── quality/                  # re-runnable QA proof (split_disjointness.py)

Audio location: only tts-premium carries actual (embedded, playable) audio. clips/segments are metadata — their file_name/audio_file describe the source mp3 layout and those raw mp3s are not bundled in this repo. For ready audio use tts-premium; to rebuild the full multi-reader audio, fetch source mp3s via clips["audio_url"]. · Sadece tts-premium gerçek (gömülü) sesi taşır; clips/segments metadata'dır, ham mp3'ler repo'da yok.

5 · Data fields

tts-premium (parquet — one row per gold clip · default, playable)

Field	Type	Description
audio	Audio (24 kHz)	embedded waveform — plays in the Viewer; decodes to array+sr
text	string	normalized transcript (TTS-ready: numbers verbalized, markup stripped)
text_raw	string	verbatim source text
book_uid	string	source book id
page	int	source page number
reader	string	reciter (A.Köseoğlu)
language	string	tur
duration	float	clip length (s)
qtier	string	quality tier: S or A (gold = S+A)
snr_db	float	SNR proxy (RMS − noise floor, dB)
split	string	row split label (train/val/test). The config is partitioned into HF splits train 40,274 / validation 1,903 / test 1,779 — leakage-safe (book/page-disjoint)

clips (clips/clips.parquet — one row per recording · metadata)

Field	Type	Description
file_name	string	original source-layout path (audio/{book}/...mp3) — raw mp3 not bundled here
audio_url	string	original source URL (rnknesriyat.com) — fetch raw audio from here
book_uid / book_name	string	book id / display name
clip_id	int	clip index within the book
reader_name / reader_id	string / int	reciter
section_no / section / subsection	int / string / string	structural location (nullable sub-fields)
description	string	optional note (often null)
page_first / page_last	int	source page range
duration_sec	int	clip duration (s)
file_size	int	bytes (verified vs server Content-Length)
text_available	bool	whether full source text exists for the book
alignment_file	string	path to the alignment .jsonl, or null
training_grade	bool	true for the 15 text books; false for the 6 text-less (Qur'an / evrad / page-scans / Kurdish / Müstakil) — machine-visible "do not build Qur'an-TTS" flag

segments (segments/segments.parquet — one row per aligned span · metadata)

Field	Type	Description
audio_file	string	the source mp3 a span belongs inside (not bundled here)
book_uid / reader_name	string	book / reciter
segment_index	int	span order within the clip
page	int	source page number
language	string	tur · ara · fas (ISO-639-3)
status	string	matched (training-grade, has timestamps) or unmatched
start / end	float	span offsets in audio_file, seconds — null when unmatched
text	string	aligned transcript

matched vs unmatched — matched (66,616) spans have start/end → usable as ASR/TTS labels. unmatched (40,787) carry text + page + language but no timestamps — text the aligner could not place. Filter status == "matched".

page-text (page-text/clip_page_text.parquet — one row per clip · clip-level source page text · metadata)

Field	Type	Description
clip_id / file_name / canonical_audio_id	int / string	clip id / path / basename (co-locates cross-book duplicate audio)
audio_url	string	original source URL (rnknesriyat.com)
duration_sec	int	clip duration (s)
book_uid / book_name / reader / reader_id	string/int	book / reciter
section / subsection	string	structural location
page_first / page_last / n_pages	int	source page range covered
page_offsets	list[[page, char]]	char offset of each page start within page_text (localize a phrase to a page)
n_missing	int	count of source pages absent in the .txt (front/back-matter; only 4 clips)
has_alignment / n_matched_seg	bool / int	whether the clip also has segment-level alignment
page_text	string	concatenated source text of the page range
n_chars / n_words	int	length
leakage_group	int	split group — clips sharing pages / identical text / audio basename share it (402 groups)

page-index (page-index/page_index.parquet): one row per {book_uid, page} → text + n_clips_covering (5,938 pages).

Value: page-resolution source text for all 1,422 text-book clips, incl. 631 with NO segment alignment (22+ reciters) — fills the speaker-diversity gap in the single-voice-dominant segments/tts-premium. Verified byte-faithful (median segment⊆page_text token-containment 1.000 under NFC; 0 join errors). Use for weak-supervision/pseudo-label ASR, long-form alignment seeding, read-along/page-sync, multi-reader eval. ⚠️ Leakage-safe split (mandatory): split with GroupKFold on leakage_group (+ reader-disjoint test) — never random or book-level split (the same passages are read by many reciters).

6 · Quick start

pip install -U "datasets>=2.19,<4.0" soundfile
# public dataset — no login required to download
# datasets<4 auto-decodes audio via soundfile; on datasets>=5 also: pip install torchcodec
#   (or use the torchcodec-free soundfile snippet below)

from datasets import load_dataset
REPO = "risaleinur/risale-nur-audio"

# ▶️ Playable audio (default config): gold single-voice TTS clips, audio embedded
tts = load_dataset(REPO, "tts-premium")     # splits: train 40,274 / validation 1,903 / test 1,779
row = tts["train"][0]
print(row["text"], "|", row["qtier"], row["snr_db"], "dB")
wav, sr = row["audio"]["array"], row["audio"]["sampling_rate"]   # 24 kHz mono

# torchcodec-free: decode the embedded bytes yourself with soundfile
from datasets import Audio
import soundfile as sf, io
tts = load_dataset(REPO, "tts-premium", split="train").cast_column("audio", Audio(decode=False))
wav, sr = sf.read(io.BytesIO(tts[0]["audio"]["bytes"]))         # 24 kHz mono

# Quality / subset filters on the playable set
gold_S  = tts.filter(lambda r: r["qtier"] == "S")              # top tier
by_book = tts.filter(lambda r: r["book_uid"] == "sozler")

# Metadata configs (no audio bundled): full catalogue + all alignment spans
clips    = load_dataset(REPO, "clips",    split="train")       # 2,429 rows (+ training_grade, audio_url)
segments = load_dataset(REPO, "segments", split="train")       # 107,403 aligned spans
matched  = segments.filter(lambda r: r["status"] == "matched") # 66,616 (text+page+timestamps)
# To rebuild full multi-reader audio, download source mp3s via clips["audio_url"],
# then slice each by segments' start/end. The raw mp3 corpus is NOT in this repo.

# Page-resolution source text for EVERY clip incl. unaligned readers (weak-supervision / read-along)
pt    = load_dataset(REPO, "page-text",  split="train")   # 1,422 clips (631 unaligned have text too)
pages = load_dataset(REPO, "page-index", split="train")   # 5,938 pages → source text
# LEAKAGE-SAFE split: GroupKFold on r["leakage_group"] (+ reader-disjoint) — never random/book split
from collections import defaultdict
groups = defaultdict(list)
for i, r in enumerate(pt): groups[r["leakage_group"]].append(i)   # keep each group in ONE fold

7 · Provenance, quality & coverage

• Provenance: source from RNK Neşriyat (rnknesriyat.com), included with permission; prepared/maintained by Şahsı Manevi (not an official RNK product). Every source mp3 was byte-verified vs server Content-Length (2,429/2,429; zero missing).
• tts-premium build: matched A.Köseoğlu Turkish spans → edge-silence trim + EBU R128 loudness + mono 24 kHz FLAC + text normalization + dedup + leakage-safe split, then a quality gate (S+A of S/A/B/C tiers).
• Alignment: forced alignment produced 791 files → 107,403 spans; matched (66,616) carry timestamps. The aligned audio is overwhelmingly one studio-grade voice, A.Köseoğlu (~96.5% of matched audio).

tts-premium quality tiers (from qtier):

Tier	Clips	~Hours
S	32,665	79.2
A	11,291	23.8
Gold (S+A) = this config	43,956	~103

📚 Per-book breakdown (corpus: clips · hours · aligned · matched segments)

Book	Clips	Hours	Aligned clips	Matched seg	Matched h	Text
Sözler	261	58.2	121	8,718	18.2	✅
Şualar	171	47.4	101	6,091	12.4	✅
Tarihçe-i Hayat	147	44.9	86	9,125	19.1	✅
Mektubat	168	41.7	80	6,314	12.5	✅
Lem'alar	122	33.5	77	4,494	9.5	✅
Emirdağ Lahikası 1	71	24.0	33	3,856	7.9	✅
İşaratü'l-İ'caz	81	23.3	42	3,290	5.7	✅
Kuran-ı Kerim	114	22.6	0	0	0	❌
Mesnevi-i Nuriye	68	21.8	27	3,757	7.9	✅
Kur'an (sayfa sayfa)	605	21.2	0	0	0	❌
Asâ-yı Mûsa	75	19.7	47	2,593	5.6	✅
Barla Lahikası	57	19.6	40	4,728	9.3	✅
Emirdağ Lahikası 2	56	18.8	28	3,193	6.5	✅
Sikke-i Tasdik-i Gaybî	51	17.4	27	3,645	6.5	✅
Kastamonu Lahikası	39	13.7	27	2,641	5.3	✅
Müstakil Risaleler	28	9.7	0	0	0	❌
İman ve Küfür Muvazeneleri	33	7.6	33	2,933	5.9	✅
Muhakemat	22	6.1	22	1,238	3.3	✅
Hizbü'l-Hakaik, Evrad	21	5.0	0	0	0	❌
H.Hakaik (sayfa sayfa)	237	4.8	0	0	0	❌
Diğer Lisanlar (Kurdish)	2	1.1	0	0	0	❌

The six text-less books are recitation / page-scan / Kurdish audio, with no aligned text by design.

🎙️ Per-reciter breakdown (corpus, top reciters)

Reciter	Clips	Hours	Matched seg	Content
A.Köseoğlu	764	~225 (≈130.9 aligned)	64,843	Risale-i Nur — single-voice TTS core
İmam Mahir	719	43.7	—	Hizbü'l-Hakaik + Qur'an (Arabic)
N.Kaya	181	41.5	—	Risale-i Nur
S.Gündoğdu	126	36.0	—	Risale-i Nur
O.Çalık	116	24.8	78	Risale-i Nur
M.Akçay	36	11.1	1,695	Risale-i Nur (+ Kurdish)
… + 20 more reciters

Only three reciters appear in the aligned (matched) set; for general Turkish ASR, balance the single-voice core with the full 462 h / 26-reciter pool to avoid single-speaker overfit.

---

7b · Reproducibility & integrity

Built deterministically (no manual relabel). Single source of truth for every number: manifests/stats.json; integrity: manifests/CHECKSUMS.sha256 (sha256 of every shipped file). Split-safety proof: quality/split_disjointness.py → train/validation/test are (book,page)-disjoint (verified 0 overlap). Full datasheet: DATASHEET.md.

• Clean (build/clean_tts.py): mono 24 kHz FLAC · loudnorm I=-23:TP=-1.5:LRA=11 · edge-trim silenceremove …-45dB:0.08s (pad 0.20 s) · dur 1–20 s · text NFC + markup/footnote strip + num2words(tr) · text-dedup.
• Split: md5("{book_uid}:{page}")%100 → <5 test, <10 validation, else train (deterministic, book/page-disjoint).
• Quality (build/score_quality.py, ffmpeg astats): snr_db = min(60, RMS−noise_floor); tier S≥38 / A 30–38 / B 22–30 / C <22 dB or clip (peak≥−0.1 dB or flat>1.0). tts-premium = S+A.
• Page map (build/build_page_map.py): source page text from metinler/{book}.txt #N markers; byte-faithful (median segment⊆page-text containment 1.000 NFC; 0 join errors).
• QA proofs ship in quality/; the build pipeline (ffmpeg + num2words + datasets/pyarrow) is documented here for reproducibility (build scripts themselves are not bundled).

8 · License & use

The source content is from RNK Neşriyat (rnknesriyat.com) and is included with their written permission (June 2026). Prepared and maintained by Şahsı Manevi; not an official RNK product. license: other — free for non-commercial use.

You may freely download, use and share this dataset for research, education, accessibility and other non-commercial purposes, provided you:

• Attribute the source — "Risale-i Nur audio — RNK Neşriyat, used with permission."
• Keep it non-commercial. Commercial use — paid products, API services, ad-supported platforms, subscriptions, closed commercial models, or any directly revenue-generating use — needs separate written permission from RNK Neşriyat.
• Don't clone or imitate a reciter's voice, or present synthetic audio as a real reciter — that needs the reciter's own consent (RNK's permission covers the recordings, not a voice-likeness license).
• Don't generate synthetic Qur'an recitation — the Qur'an audio here is authentic human recitation and must stay so.
• Treat the printed text as authoritative for wording; recordings may contain rare reading or technical slips.

Corrections and takedown requests from rights-holders (RNK, reciters, their representatives) are honored in good faith via the Community tab. See NOTICE.md.

🇹🇷 Türkçe

Kaynak içerik RNK Neşriyat'tandır (rnknesriyat.com) ve yazılı izinleriyle (Haziran 2026) yer alır. Şahsı Manevi tarafından hazırlanır; RNK'nın resmî ürünü değildir. license: other — ticari olmayan kullanıma açık.

Veri kümesini araştırma, eğitim, erişilebilirlik ve diğer ticari olmayan amaçlarla serbestçe indirip kullanabilir ve paylaşabilirsiniz; yeter ki:

• Kaynağı belirtin — "Risale-i Nur sesleri — RNK Neşriyat, izinle."
• Kullanım ticari olmasın. Ticari kullanım (ücretli ürün, API servisi, reklamlı platform, abonelik, kapalı ticari model veya doğrudan gelir getiren her kullanım) RNK Neşriyat'tan ayrı yazılı izin gerektirir.
• Okuyucu sesini klonlamayın/taklit etmeyin, yapay sesi gerçek okuyucu gibi sunmayın — bu, okuyucunun rızasını gerektirir.
• Yapay Kur'an tilaveti üretmeyin — buradaki Kur'an sesi gerçek beşerî tilavettir, öyle kalmalıdır.
• Lafız için basılı metni esas alın; kayıtlarda nadiren okuma/teknik kusur olabilir.

Hak sahiplerinin (RNK, okuyucular, temsilcileri) düzeltme/kaldırma talepleri Community sekmesinden iyi niyetle karşılanır.

9 · Intended use & ethics

✅ In scope — Turkish/Arabic ASR · non-Qur'anic TTS (incl. accessibility read-aloud) · forced-alignment & linguistics · audio classification · Islamic-studies / digital-humanities research.

🚫 Out of scope — synthetic Qur'an recitation; cloning or impersonating a reciter's voice without consent; commercial use without separate RNK permission (see §8 · License & use).

⚖️ Note — religiously sensitive material; please use it respectfully (emanet).

Limitations & biases — (a) the aligned / tts-premium core is single-speaker (A.Köseoğlu ≈97% of matched) → well-suited to single-speaker TTS research, weak for general Turkish ASR (balance with the 462 h / 26-reciter pool); (b) Arabic / Persian / Kurdish are under-represented; (c) the 40,787 unmatched segments and the 631 unaligned page-text clips have no timestamps; (d) page_text retains footnote (hâşiye) prose that may not be spoken.

🇹🇷 Türkçe

✅ Kapsam — Türkçe/Arapça ASR · Kur'an dışı TTS (erişilebilirlik dahil) · hizalama/dilbilim · ses sınıflandırma · İslami/dijital beşeri bilimler. 🚫 Lütfen yapmayın — yapay Kur'an tilaveti üretmek; yapay sesi gerçek tilavet gibi sunmak; bir okuyucunun sesini rızası olmadan klonlamak/taklit etmek (RNK izni kayıtları kapsar, ses-benzerliği lisansı değil). ⚖️ Not — dinen hassas materyal; hürmetle kullanın. Sınırlar: çekirdek tek-konuşmacı (genel ASR için 26-okuyucu havuzuyla dengele); Arapça/Farsça/Kürtçe az temsilli; unmatched + hizalanmamış page-text kliplerinin zaman damgası yok.

---

10 · Maintainers & dataset family

Prepared and maintained by Şahsı Manevi, with gratitude to RNK Neşriyat and to the 26 reciters whose voices make this corpus possible. Audio member of the Şahsı Manevi dataset family:

Modality	Dataset	Role
🔊 Audio (this)	risaleinur/risale-nur-audio	Spoken readings + recitation (ASR / TTS)
📖 Text	risaleinur/risale-nur-grounded-multipool	Grounded text corpus for retrieval / NLP

The two share book_uid and page keys, so audio spans can be joined to source text at page resolution.

11 · Citation

@misc{sahsimanevi2026risalenuraudio,
  title        = {Risale-i Nur Sesli K\"ulliyat: A Turkish--Arabic Speech Corpus
                  of Said Nursi's Works with Qur'an and Hizb\"ul-Hakaik Recitation},
  author       = {{\c{S}ahs{\i} Manevi}},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/risaleinur/risale-nur-audio}},
  note         = {Source: RNK Ne\c{s}riyat (rnknesriyat.com), used with permission; non-commercial. See NOTICE.}
}

Plain text: Risale-i Nur Audio Corpus (Şahsı Manevi, 2026), Hugging Face: risaleinur/risale-nur-audio. Source: RNK Neşriyat.

12 · Versioning

• v1.0 (2026-06) — initial release. tts-premium 43,956 clips (~103 h, embedded audio) · clips 2,429 · segments 107,403 (66,616 matched).
• v1.1 (2026-06) — added page-text (1,422 clips → source page text, incl. 631 unaligned + leakage_group) and page-index (5,938 pages) metadata configs.
• v1.2 (2026-06) — public release under RNK Neşriyat's written permission (non-commercial); access opened, dataset card streamlined.
• Questions or corrections? Use the Community tab. · Soru/düzeltme için Community sekmesi.