Onay otomatiktir: şartları kabul edince indirme başlar. İki yükümlülük vardır — bu veri kümesine atıf vermek ve bu veriyle eğittiğiniz modelin ağırlıklarını kamuya açık yayınlamak. Gerisi, ses dahil yeniden dağıtım ve ticari kullanım, serbesttir.

Approval is automatic: accept the terms and the download starts. There are two obligations — cite this dataset, and publish the weights of any model you train on it. Everything else, including redistribution with the audio and commercial use, is allowed.

Kullanım şartları / Terms of use

Bu veriyle ürettiğiniz her şeyde — model, türev veri kümesi, makale, ürün — bu veri kümesine atıf verirsiniz. Hub'da model yayınlıyorsanız model kartının YAML başlığına `datasets:

serdarcaglar/turkish-tts-audiobookssatırını eklersiniz. / You will cite this dataset in anything you build on it. If you publish a model on the Hub, adddatasets:
serdarcaglar/turkish-tts-audiobooks` to the model card's YAML header.

Bu veriyle eğittiğiniz her modelin ağırlıklarını, ticari kullanıma ve yeniden dağıtıma izin veren bir lisansla (Apache-2.0, MIT, CC-BY vb.), modeli ekibiniz dışına ilk sunduğunuz günden geç olmamak üzere kamuya açık yayınlarsınız. Kapalı ağırlıklı veya yalnızca API arkasından sunulan model eğitmek yasaktır. / You will publish the weights of any model you train on this data under a license that permits commercial use and redistribution (Apache-2.0, MIT, CC-BY or equivalent), no later than the day you first make the model available outside your team. Closed-weight or API-only models are not permitted.
Ağırlıkları yayınladıktan sonra modeli ticari olarak kullanabilirsiniz; sesi ve transcript'leri yeniden dağıtabilir, ses dahil türev veri kümesi yayımlayabilirsiniz. Tek koşul, bu iki yükümlülüğü kendi kullanıcılarınıza da aktarmanızdır. / Once the weights are public you may use the model commercially, redistribute the audio and transcripts, and publish derived corpora with the audio included — provided you pass these two obligations on to whoever uses your version.
Kaynak kayıtların hakları temizlenmemiştir; hukuki sorumluluğu üstlenir ve hak sahibi talebi üzerine kopyalarınızı silersiniz. Ayrıca kimsenin sesini rızası olmadan klonlamamanızı rica ederiz. / The rights of the source recordings are not cleared; you accept legal responsibility for your use and will delete your copies upon a rightsholder request. We also ask that you not clone anyone's voice without their consent.

Turkish TTS Audiobooks

Turkish read-speech corpus for text-to-speech training, built from Turkish audiobook and spoken-article recordings by an automatic pipeline: VAD segmentation → technical QC → acoustic event tagging → DNSMOS → speaker embedding/consistency → double-pass Whisper ASR → text policy → leakage-free splitting. Audio is 16 kHz mono lossless FLAC embedded in the Parquet shards.

The pipeline that produced it — every stage, every threshold, the export and audit scripts — is open source at serdarildercaglar/turkish-tts-audiobooks.

857,123 clips · 2,724 hours · 16 kHz mono · machine transcripts.

The corpus ships in two pools, and the difference matters more than anything else on this page:

train + validation — 424,817 clips, 1,311 hours. Every clip passed every filter in Filtering policy. This is the part you can train on directly.
review — 432,306 clips, 1,413 hours. Clips that tripped at least one review threshold. Not rejects — the pipeline's hard failures were thrown away and never uploaded — but unverified. 83% of the pool carries the suspected-synthetic-speech flag, from a check that is deliberately paranoid and also fires on clean studio narration: in a 240-clip blind listening audit (10 per channel), every sampled clip — including all 131 flagged ones — was judged human narration. Still: mine it, don't dump it into your training set. See The review split.

Türkçe özet

Türkçe sesli kitap ve sesli makale kayıtlarından otomatik olarak üretilmiş TTS korpusu: 857.123 klip, 2.724 saat, 16 kHz mono kayıpsız FLAC. Korpusu üreten hattın tamamı açık kaynak: serdarildercaglar/turkish-tts-audiobooks. Transcript'ler whisper-large-v3-turbo ile üretildi ve ikinci bir geçişle karşılaştırıldı; klip başına insan doğrulaması yoktur (180 kliplik insan referanslı örneklem denetiminde temiz havuz CER'i 0,0012 ölçüldü). Metinde rakamlar, noktalama ve kesme işaretleri kaynaktaki gibi korunur, yazıya çevrilmez ("2024'te" → "2024'te").

İki havuz var: train+validation (424.817 klip / 1.311 saat) tüm filtrelerden geçti, doğrudan eğitime uygundur. review (432.306 klip / 1.413 saat) ise filtrelerin emin olamadığı kliplerdir — çöp değildir, kesin elenenler zaten yüklenmedi — ama doğrulanmamıştır; havuzun %83'ü "sentetik ses şüphesi" işareti taşıyor ve bu kontrol temiz stüdyo anlatımını da yakalıyor: 240 kliplik kör dinleme denetiminde işaretli 131 klip dahil örneklenen her klip insan anlatımı çıktı. Kontrol etmeden eğitime katmayın. Karar politikasının okuduğu ölçümlerin tamamı yayımlanır: yedisi satırın içinde, geri kalanı metrics/quality_metrics_full.parquet yan dosyasında, klip idsiyle eşlenmiş olarak.

Kaynak kayıtların telif durumu doğrulanmamıştır. Erişim forma tabidir ama onay otomatiktir. İki şart var: bu veri kümesine atıf vermek, ve bu veriyle eğittiğiniz modeli açık ağırlıkla yayınlamak — kapalı ağırlıklı ya da yalnızca API arkasından sunulan model eğitmek yasaktır. Ağırlıkları yayınladıktan sonra modeli ticari olarak kullanmakta serbestsiniz. Bu şartlar başka kısıt getirmez; ancak bu yayın size kaynak ses üzerinde hak devretmez — yeniden dağıtım dahil her kullanımın hukuka uygunluğunu kendi hukukunuzda değerlendirmek size düşer (bkz. Access and terms, Licensing).

Splits

Split	Clips	Hours	Sources	Channels	Speaker clusters	Clips with `ref_id`	Use it for
`train`	416,315	1,292.2	1,179	23	744	143,854 (34.6%)	training
`validation`	8,502	19.0	915	22	384	1,604 (18.9%)	held-out eval
`review`	432,306	1,412.7	2,324	24	—	—	mining, after your own filtering

train/validation are split at source level: all clips of one recording land in the same split, so no recording — and therefore no book or session — crosses the boundary. Validation is filled by taking the smallest sources (seed 42) until 2% of clips is reached, so it is leakage-free at the audio level but not distribution-matched: validation clips are shorter (median 7.0 s vs 11.0 s) and louder (median −18.6 LUFS vs −22.7 LUFS) than training clips, and one small channel appears only there.

Treat validation as a smoke test for narrator overfitting, not as a comparative benchmark. Besides the distribution mismatch, 501 of its 8,502 clips (5.9%) share verbatim normalised transcript text with train — the same book read by a different narrator on a different channel, which source-level splitting cannot see. If you need a proper held-out set, carve one from train by source_id and filter transcript overlap out.

review is not disjoint from train: the same recording can contribute accepted clips to train and flagged clips to review. Never evaluate a train-trained model on review, and keep source_id groups intact if you resplit anything.

Which split do I want?

If you want to…	Use
fine-tune a TTS model today	`train`, optionally filtered harder on `quality_*`
measure held-out quality	`validation` (or carve your own from `train` by `source_id`)
zero-shot / voice-prompt training pairs	`train` rows where `ref_id` is set
more speakers, more hours, and you can spend effort verifying	`review`, filtered by `review_reasons` — see the recipe
a clean benchmark	none of these as-is: transcripts are machine-generated

Fields

Field	Type	Description
`id`	string	Unique clip id (`<source_id>-<start_ms>-<end_ms>`)
`audio`	Audio(16 kHz)	Mono lossless FLAC, embedded in the Parquet shard
`text`	string	Training transcript (conservative cleanup)
`text_raw`	string	Unmodified ASR output
`text_normalized`	string	Current training text; byte-identical to `text` in v1.0.0, kept as a stable slot for future normalization
`duration`	float32	Seconds (2.5–20.0 clean, up to 20.4 in `review`)
`sample_rate`	int32	16000 for every row
`language`	string	`tr`
`speaker_id`	string	`<channel>-<NNN>`, channel-local embedding cluster. Empty in `review`
`channel`	string	Source collection/uploader the recording came from
`source_id`	string	Stable, non-path recording id; grouping key for resplitting
`source_start_sec`, `source_end_sec`	float32	Clip location inside the original recording
`ref_id`	string	Optional id of a voice-prompt clip for zero-shot TTS: same split and channel, 3–10 s long, embedding cosine ≥ 0.80, always from a different source recording. Audio is not duplicated; resolve by `id`. Empty in `review`
`split`	string	`train` / `validation` / `review`
`decision`	string	`ACCEPT` in `train`/`validation`, `REVIEW` in `review`
`dataset_version`	string	`1.0.0`
`transcription_model`	string	`openai/whisper-large-v3-turbo`
`license`	string	Per-row license slot; `other` (see Licensing)
`review_reasons`	list[string]	Empty for accepted clips; in `review`, the thresholds the clip tripped
`quality_speech_ratio`	float32	VAD speech seconds / clip seconds
`quality_clip_ratio`	float32	Fraction of samples at digital full scale
`quality_lufs`	float32	Integrated loudness, not normalized
`quality_music_score`	float32	AudioSet music-label score (max over windows)
`quality_dnsmos_ovrl`	float32	DNSMOS P.835 overall MOS
`quality_speaker_cosine`	float32	Cosine to the robust embedding centroid of the same source recording (intra-recording voice consistency)
`quality_asr_cer`	float32	CER between the two independent ASR passes (transcript disagreement proxy)

Full per-clip metrics (sidecar)

The seven quality_* columns above are the metrics that drove the accept/review decision and are stored in-row. Every other measurement the pipeline read is also published, in one sidecar file keyed by id: metrics/quality_metrics_full.parquet — 857,123 rows (one per published clip, all three splits), ~51 MB.

Sidecar column	Description
`id`, `split`	Join keys — same values as the dataset rows
`synthetic_speech_score`	AudioSet "Speech synthesizer" score; the signal behind `sentetik_ses_suphesi` (83% of the review pool)
`dnsmos_sig`, `dnsmos_bak`	DNSMOS P.835 signal and background MOS (OVRL is in-row)
`audioset_speech_score`	AudioSet "Speech" label score
`internal_silence_sec`	Longest internal silence
`boundary_risk`	Forced cut may fall inside a word (`kesim_siniri_riski`)
`rms_db`, `dc_offset`, `peak_db`	Signal-domain QC
`event_score`	Top non-speech AudioSet event score
`max_ngram_repeat`, `asr_avg_logprob`, `asr_no_speech_prob`	ASR decoder diagnostics
`n_words`, `n_chars`, `words_per_sec`, `foreign_ratio`	Text statistics

Together the in-row columns and the sidecar contain all the inputs of the decision policy, so any stricter or looser cut — including a different synthetic-speech threshold — is reproducible without rerunning any model:

import pandas as pd
from huggingface_hub import hf_hub_download

metrics = pd.read_parquet(hf_hub_download(
    "serdarcaglar/turkish-tts-audiobooks", "metrics/quality_metrics_full.parquet", repo_type="dataset"))
synth = dict(zip(metrics["id"], metrics["synthetic_speech_score"]))

# e.g. re-cut the synthetic-speech threshold at 0.60 instead of 0.45
recovered = review.filter(lambda r: synth[r["id"]] < 0.60)

Statistics

Median values, and the range you actually get. train on the left, review on the right — the gap is the whole story of the two pools.

Metric	train p05	train median	train p95	review p05	review median	review p95
duration (s)	4.75	11.03	16.78	6.70	11.31	17.71
DNSMOS OVRL	3.10	3.36	3.53	2.95	3.36	3.53
speech ratio	0.82	0.92	0.98	0.78	0.92	0.98
speaker cosine	0.84	0.92	0.95	0.82	0.92	0.96
ASR CER (2-pass)	0.000	0.000	0.018	0.000	0.000	0.019
music score	0.001	0.002	0.015	0.000	0.001	0.561

Worst case per pool (min/max), which is what the filters are actually about:

Metric	train	review
DNSMOS OVRL	2.70 – 3.71	1.81 – 3.77
speech ratio	0.75 – 1.00	0.18 – 1.00
speaker cosine	0.55 – 0.98	−0.11 – 0.99
ASR CER	0.00 – 0.15	0.00 – 49.5
music score	0.000 – 0.250	0.000 – 0.700
loudness (LUFS)	−40.0 – −10.9	not bounded below −40

Other train figures: text length median 148 chars (max 374), speaking rate median 1.83 words/s (0.80–3.70), loudness median −22.7 LUFS (p05–p95: −28.5 to −16.1). Duration is bimodal by design — the segmenter targets ~9 s and splits long speech runs, so 61% of clips fall in the 9–13 s band (69% within 9–14 s).

Human-reference transcript audit

A 180-clip fixed-seed sample (120 clean + 60 review, spread across source recordings, 23 of 24 channels) was listened to and corrected to verbatim human references by one annotator. 170 of 180 transcripts (94.4%) needed no edit. Micro-averaged CER of the released text against the corrected references: 0.0012 on the clean sample, 0.0193 on the review sample (driven by two clips at 0.69 and 0.48), 0.0080 overall. A sample-level average, not a per-clip guarantee — the corpus is still not an ASR benchmark. The sampling/correction harness is in the pipeline repo under audit/.

Channel distribution

Clips per channel, all three splits. Note how many channels live almost entirely in review — those narrators are missing from train.

Channel	train	validation	review
seslikitaplarmavi	172,933	149	15,975
BirDinle	158,884	94	9,981
anahtarca	21,117	227	11,455
ZubeyirSener	13,296	134	47,748
dinleyiniz	7,973	384	65,437
OkumaSaati	7,202	0	560
seslimakalem	6,358	802	2,641
seslikutuphanemkanali	4,870	370	16,113
kitaplar	4,847	964	28,565
sess-Seslikitap	4,294	544	65,123
eba	3,156	240	4,432
ses-arşiv	2,299	1,281	45,591
bizimkütüphane	1,844	0	47
Seslendiriyor	1,484	242	6,218
SESLİKİTAPEVİ	1,407	191	5,899
idea_stüdyo	1,333	269	16,192
sesli-kitaplar	1,047	168	5,967
cantadakitap	758	1,138	33,651
MuratKaraOfficial2021	724	922	12,273
seslikitapturkish	367	81	9,268
denizinötesindekisesler	47	136	4,369
KitaplarinKedisi	42	34	264
seskitap	33	80	2,682
kitapdinle	0	52	21,855

Two channels carry 80% of the training hours, so weight or subsample by channel / speaker_id if balanced speaker coverage matters. The reverse is also true: dinleyiniz, sess-Seslikitap, ZubeyirSener, ses-arşiv, cantadakitap and kitapdinle are effectively review-only.

Provenance

Every setting below lives in one file, configs/default.yaml, and the stages that apply them are in datapipe/.

Step	Model / setting
Decode & resample	ffmpeg → 16 kHz mono, 40 Hz high-pass, peak ceiling −1 dB, no per-clip loudness normalization
Segmentation	Silero VAD (threshold 0.50), min speech 250 ms, min silence 300 ms, 100 ms pad, target 9 s, hard bounds 2.5–20 s
Technical QC	duration / clipping / RMS / DC offset / internal silence / speech ratio
Acoustic events	`MIT/ast-finetuned-audioset-10-10-0.4593`, 10.24 s windows, 3 s hop — music, noise events, synthetic-speech labels
Perceptual quality	DNSMOS P.835 (`sig_bak_ovr.onnx`)
Speaker embedding	`pyannote/wespeaker-voxceleb-resnet34-LM`, L2-normalized
Speaker consistency	cosine to a trimmed-mean centroid per source recording
ASR	`openai/whisper-large-v3-turbo` on vLLM, `language=tr`, two independent passes (greedy + temperature 0.4); CER between passes is the transcript-reliability signal
Text policy	length, words/second, foreign-character ratio, n-gram repetition
Speaker clustering	greedy online nearest-centroid over embeddings, cosine 0.75, ≤64 clusters per channel (accepted clips only)
Reference pairing	within (split, channel), cross-recording, 3–10 s, cosine ≥ 0.80, ≤512-token budget, ~40% of clips targeted (accepted clips only)

Text conventions

Transcripts are ASR output with conservative cleanup only. Digits, punctuation and Turkish apostrophes are kept verbatim — 2024'te stays 2024'te, no number/abbreviation expansion, no case folding, no punctuation stripping. If your frontend needs verbalized numbers, normalize downstream. text differs from text_raw in roughly 4% of clips (whitespace, stray symbols, dangling fragments); text_normalized is byte-identical to text in this version.

Filtering policy

Every measurement is per clip. Three outcomes: clean (train/validation), flagged (review, published with its reasons), hard reject (deleted, not in this dataset).

Signal	Clean requires	Flagged → `review`	Hard reject
duration	2.0–21.0 s	— (see note)	< 2.0 s or > 21.0 s
clipping ratio	≤ 0.001	0.001–0.02	> 0.02
speech ratio (VAD)	≥ 0.75	< 0.75	—
internal silence	≤ 1.0 s	> 1.0 s	—
loudness / level	≥ −40 LUFS, RMS ≥ −50 dB	below that	RMS < −60 dB
DC offset	< 0.01	≥ 0.01	—
music score	≤ 0.25	0.25–0.70	> 0.70
other acoustic events	≤ 0.35	0.35–0.75	> 0.75
AudioSet speech score	≥ 0.30	< 0.30	—
synthetic-speech score	< 0.45	≥ 0.45	—
DNSMOS	sig ≥ 3.0, bak ≥ 3.0, ovrl ≥ 2.7	between the bars	sig/bak < 2.0, ovrl < 1.8
speaker cosine (in-recording)	≥ 0.55	< 0.55	—
2-pass ASR CER	≤ 0.15	> 0.15	n-gram repeat ≥ 5
text	≥ 2 words, ≤ 400 chars, 0.8–6.5 words/s, foreign chars ≤ 5%	outside those bounds	< 2 words

Note on duration: the segmenter only emits clips of 2.5–20.4 s, so the duration bars never fire in this release and duration produces no review reason. Clips whose forced cut may fall inside a word carry the separate kesim_siniri_riski flag instead.

The seven decision-driving metrics are stored in-row; every remaining policy input is in metrics/quality_metrics_full.parquet, keyed by id (see Full per-clip metrics). Between the two, you can reproduce a stricter or looser cut of every threshold above without rerunning the pipeline.

Yield

2,370 source recordings (19 contained no speech, 1 failed to decode) produced 879,049 candidate clips:

Outcome	Clips	Hours	In this dataset
ACCEPT	424,817	1,311.2	yes — `train` + `validation`
REVIEW	432,306	1,412.7	yes — `review`, unverified
REJECT	21,926	66.9	no

The review split

review holds the clips the automatic filters were not confident about. They are published rather than discarded so the call can be made by a human — or by a better classifier — instead of the data being lost. Hard failures never made it here; those were rejected outright.

Do not concatenate review onto train and start a run. Pick the reasons you can live with, or re-filter with your own model.

Why clips were flagged

One clip can carry several reasons: 347,958 have exactly one, 84,348 have two or more (up to nine).

Reason (`review_reasons`)	Clips	What it means
`sentetik_ses_suphesi`	358,934	AudioSet "Speech synthesizer" score ≥ 0.45 — suspected TTS narration
`sinirda_muzik`	63,220	music score 0.25–0.70 (background bed, intro/outro)
`clipping`	56,935	0.1–2% of samples at full scale
`uzun_ic_sessizlik`	19,463	internal silence > 1.0 s
`dusuk_konusma_orani`	13,750	VAD speech ratio < 0.75
`sinirda_dnsmos_ovrl` / `_bak` / `_sig`	5,192 / 4,364 / 567	DNSMOS below the clean bar but above hard reject
`kesim_siniri_riski`	3,314	clip boundary may cut a word
`asr_cift_gecis_uyusmazligi`	1,961	the two ASR passes disagree (CER > 0.15)
`farkli_konusmaci_suphesi`	1,515	embedding cosine to the recording centroid < 0.55
`konusma_hizi`	1,300	outside 0.8–6.5 words/second
`ast_dusuk_speech`	702	AudioSet speech score < 0.30
`dusuk_lufs`	571	integrated loudness < −40 LUFS
`ses_olayi`	293	non-speech event (applause, vehicle, phone…) score 0.35–0.75
`uzun_metin` / `yabanci_karakter`	62 / 29	transcript over 400 chars / >5% non-Turkish characters

83% of the pool is the synthetic-speech flag. That check is deliberately paranoid: the AudioSet label also fires on close-mic, compressed, evenly-paced studio narration — exactly what a good audiobook sounds like. A blind listening audit put a number on it: one annotator judged 240 clips (10 per channel, channel and score hidden) and found all 240 human, including every one of the 131 flagged clips — sample precision 0.00, and all 24 channels, kitapdinle (flag rate 0.998) included, judged human on 10/10. On this evidence the flag detects studio production, not synthesis, and the ~1,200 flagged hours are recoverable human narration. Caveats: one listener, 10 clips per channel, and only audible synthesis is bounded — verify per channel before you train.

How `review` differs from the clean splits

No speaker_id, no ref_id — speaker clustering and reference pairing run on accepted clips only. Both fields are empty here.
Sources overlap train, so review is unusable as an evaluation set for a model trained on train.
Wider quality range — see the min/max table in Statistics. The CER outlier (up to 49.5) marks ASR repetition loops; those transcripts are garbage and easy to filter.
Different channel mix — six channels are effectively review-only, including kitapdinle, which has no training clips at all. If a whole channel was flagged as suspected TTS, its speakers are missing from train entirely. This is where the pool is worth the most.

Suggested use

review = load_dataset("serdarcaglar/turkish-tts-audiobooks", split="review")

# Only the synthetic-speech suspicion, nothing else wrong, quality bar high
candidates = review.filter(
    lambda r: r["review_reasons"] == ["sentetik_ses_suphesi"]
    and r["quality_dnsmos_ovrl"] >= 3.2
    and r["quality_asr_cer"] <= 0.05
)

# Then decide channel by channel, not clip by clip
from collections import Counter
print(Counter(candidates["channel"]))

The flag's underlying score ships per clip in the metrics sidecar as synthetic_speech_score, so you can also re-threshold it directly instead of consuming the binary reason.

Listen to a handful of clips per channel before trusting it: a human ear settles the synthetic-vs-real question in minutes per channel, and the flag is channel-correlated — whole uploaders are either TTS or not. The pipeline repository ships the blind-listening harness used for exactly this check in audit/: it draws a fixed-seed sample per channel, hides the channel and the classifier score from the listener, and scores the result at both clip and channel level.

Türkçe

review, otomatik filtrelerin emin olamadığı 432.306 klip (1.413 saat). Çöp değildir; kesin elenenler (21.926 klip) zaten bu veri kümesinde yok. Havuzun %83'ü aynı işareti taşıyor: AudioSet'in "sentetik ses" etiketi, ki bu etiket temiz stüdyo anlatımında da tetikleniyor — 240 kliplik kör dinleme denetiminde işaretli 131 klip dahil her klip insan çıktı; işaret bu korpusta sentezi değil stüdyo prodüksiyonunu yakalıyor. Bu split'te speaker_id ve ref_id yoktur, kaynaklar train ile örtüşür (bu yüzden değerlendirme kümesi olarak kullanılamaz) ve kalite aralığı geniştir. Doğrudan eğitime eklemeyin: sebep filtreleyip kanal başına örnek dinleyerek seçin. Altı kanal neredeyse tamamen bu havuzda — kitapdinle'nin train'de hiç klibi yok — asıl kazanç orada.

Usage

from datasets import load_dataset

ds = load_dataset("serdarcaglar/turkish-tts-audiobooks", split="train", streaming=True)
row = next(iter(ds))
row["audio"]["array"], row["audio"]["sampling_rate"], row["text"]

Zero-shot / voice-prompt training pairs — resolve ref_id inside the same split:

train = load_dataset("serdarcaglar/turkish-tts-audiobooks", split="train")
position = {clip_id: i for i, clip_id in enumerate(train["id"])}
pairs = ((i, position[r]) for i, r in enumerate(train["ref_id"]) if r)

A stricter, studio-grade subset of the clean pool:

clean = train.filter(
    lambda r: r["quality_dnsmos_ovrl"] >= 3.4
    and r["quality_asr_cer"] == 0.0
    and r["quality_speaker_cosine"] >= 0.90
)

Audio is 16 kHz — the common denominator of the TTS models this corpus was built for (VoxCPM2's AudioVAE input is one of them). Resample if your model expects 22.05/24/44.1 kHz.

Limitations and caveats

Human verification is sample-level only. Transcripts are machine-generated. A 180-clip human-reference audit (see Statistics) measured an average clean-pool CER of 0.0012, but that covers 0.02% of the corpus — no individual transcript is guaranteed. The 2-pass CER agreement (median 0.000) bounds ASR instability, not correctness: systematic errors both passes make — proper nouns, foreign words, numbers, omitted discourse particles — survive it.
review is sample-audited, not clip-verified. The blind listening audit de-risks only the synthetic-speech flag, and only at sample level; the other review reasons — clipping, music, internal silence, low speech ratio — are measured signal defects no audit removes. Do not treat it as a second training set without your own filtering pass.
speaker_id is a cluster, not an identity. Greedy single-pass clustering runs per channel with at most 64 clusters; once a channel hits the cap, further clips are attached to the nearest cluster regardless of threshold. One narrator can hold several ids, several narrators can share an id (most likely in the two large channels, which both hit the cap), and the same person on two channels always gets two ids.
Channel imbalance: 80% of training hours come from two channels; the review pool is dominated by a different set.
Loudness is not normalized — level varies within and across channels (train p05–p95: −28.5 to −16.1 LUFS). Normalize at load time if needed.
Read speech only: audiobooks and spoken articles. Expect narration prosody; no conversational, spontaneous or emotional-speech coverage, and no dialect/regional balance guarantees.
Validation is not distribution-matched and is not a benchmark — 5.9% of its clips share verbatim normalised text with train (same book, different narrator); see Splits for how to carve a proper held-out set.
Possible near-duplicates across sources: the same book read by the same narrator may exist in more than one channel; dedup was not attempted beyond source-level grouping.
Text may contain publisher boilerplate (intro/outro announcements, channel plugs) where the narration itself contained it.
No metadata about book titles, authors, publication years or narrator demographics is included.

Access and terms

Access is behind a form, but approval is automatic: accept the terms and the download starts. Two obligations follow you afterwards, and only two.

Cite this dataset. In anything you release that was built on this data — a model, a derived corpus, a paper, a product page. If you publish a model on the Hub, put serdarcaglar/turkish-tts-audiobooks in the datasets: field of the model card's YAML header; that one line is what makes the work discoverable from this page. See Citation for the BibTeX.
Models must be open. If you train a model on this data — from scratch, by fine-tuning, or by distillation — publish its weights publicly under a license that itself permits commercial use and redistribution (Apache-2.0, MIT, CC-BY or equivalent), no later than the day you first make the model available to anyone outside your team. Serving a model through an API, product or demo without releasing its weights is the one case this dataset forbids. Once the weights are out, sell what you like: the obligation is on the weights, not on your business.

These two obligations are conditions of access — a contract you accept with the form — not a copyright licence. We hold no copyright in the source recordings and therefore cannot grant you any rights over them: their copyright status was not cleared, and no amount of automatic processing creates redistribution rights that did not exist (see Licensing). Within that limit, these terms restrict nothing else: filtering, re-labelling, redistributing your copy, publishing a derived corpus, mining review — the terms only require that you cite this dataset and pass both obligations on to whoever uses your version, so that a model trained on a derivative is still an open model that credits the source. Whether a use that copies or republishes the underlying audio is lawful is a separate question these terms cannot answer for you: assess it in your own jurisdiction, and delete affected material on a rightsholder's request.

One request that is not a condition of access, but matters: do not clone or identify a specific person's voice without their consent.

Erişime bir form üzerinden ulaşılır ama onay otomatiktir; şartları kabul edince indirme başlar. Bundan sonrası iki yükümlülükten ibarettir. Birincisi atıf: bu veriyle ürettiğiniz her şeyde — model, türev veri kümesi, makale, ürün sayfası — bu veri kümesine referans verirsiniz; Hub'da model yayınlıyorsanız model kartının YAML başlığındaki datasets: alanına serdarcaglar/turkish-tts-audiobooks satırını eklemeniz gerekir. İkincisi açık ağırlık: bu veriyle eğittiğiniz modelin ağırlıklarını, ticari kullanıma ve yeniden dağıtıma izin veren bir lisansla (Apache-2.0, MIT, CC-BY vb.), modeli ekibiniz dışına ilk sunduğunuz günden geç olmamak üzere kamuya açık yayınlarsınız; kapalı ağırlıklı veya yalnızca API arkasından sunulan model eğitmek yasaktır. Ağırlıklar yayınlandıktan sonra ticari kullanım serbesttir.

Bu iki yükümlülük, formu kabul ederek girdiğiniz erişim sözleşmesinin koşullarıdır; bir telif lisansı değildir. Kaynak kayıtların telifi bize ait olmadığı ve doğrulanmadığı için bu yayın size kaynak ses üzerinde hiçbir hak devretmez: otomatik işleme, var olmayan bir yeniden dağıtım hakkı yaratmaz. Bu sınır içinde şartlar başka hiçbir şeyi kısıtlamaz — ayıklamak, yeniden etiketlemek, review havuzundan kendi veri kümenizi türetmek serbesttir; yeter ki atıf verin ve aynı iki yükümlülüğü kullanıcılarınıza aktarın. Sesin kopyalanmasını veya yeniden yayımını içeren bir kullanımın hukuka uygunluğunu kendi hukukunuzda siz değerlendirirsiniz ve hak sahibi talebinde ilgili materyali silersiniz. Şart değil ama ricamız: kimsenin sesini rızası olmadan klonlamayın.

Citation

If you use this dataset, or anything derived from it, cite it. Author: Serdar I. Çağlar, ORCID 0000-0002-5776-2431.

@misc{caglar2026turkishttsaudiobooks,
  title        = {Turkish TTS Audiobooks: a 2{,}724-hour Turkish read-speech
                  corpus for text-to-speech},
  author       = {Serdar I. {\c{C}}a{\u{g}}lar},
  orcid        = {0000-0002-5776-2431},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/serdarcaglar/turkish-tts-audiobooks}}
}

For a model published on the Hub, the citation that actually gets counted is one line in the model card's YAML header:

datasets:
  - serdarcaglar/turkish-tts-audiobooks

Bir modeli veya türev veri kümesini yayınlarken yukarıdaki BibTeX'i kullanın; Hub'da yayınlıyorsanız kart başlığındaki datasets: satırı da zorunludur.

Author

Serdar I. Çağlar — ORCID 0000-0002-5776-2431 · Hugging Face. Corpus construction, pipeline and dataset card.

Questions, corrections and takedown requests: open a discussion on this repository.

Licensing

license: other — there are two layers here, and they are different things.

The access terms (previous section) are a contract between you and the publisher of this dataset: cite the dataset, and release the weights of any model trained on it. They are the only conditions the publisher imposes.

The underlying rights are not the publisher's to give. The source recordings were collected in mid-2026 from publicly accessible Turkish audiobook and spoken-article upload channels; the uploader is recorded per row in channel, and each clip's position in its source recording in source_id / source_start_sec / source_end_sec. The recordings' individual copyright status was not cleared, and automatic processing grants no redistribution rights — so this release cannot and does not grant you any licence over the audio or the underlying texts. Whoever uses or publishes this dataset, a derivative, or a model trained on it is responsible for verifying the rights of the underlying recordings and transcripts in their jurisdiction. If you hold rights to material here and want it removed, open a discussion on this repository and it will be taken down.

Version

1.0.0 — produced by turkish-tts-audiobooks datapipe v2, decision thresholds as tabulated above. Regenerating with different thresholds changes clip counts; the per-clip quality_* columns and review_reasons let you reproduce a stricter or looser subset without rerunning the pipeline.

Downloads last month: 310

Total file size:

172 GB

Şartları kabul edin / Accept the terms