Kimliğinizi doğrulayabilmemiz için LinkedIn profiliniz zorunludur; formdaki ad ve kurumla uyuşmayan talepler reddedilir. / A LinkedIn profile is required so we can verify who you are; requests that do not match the name and affiliation on the form are rejected.

Kullanım şartları / Terms of use

Ticari kullanım yasaktır. Veri kümesi, türevleri veya bununla eğitilmiş modeller ücretli ürün, hizmet veya API'de kullanılamaz; satılamaz veya lisanslanamaz. / No commercial use: neither the data, its derivatives, nor models trained on it may be used in a paid product, service or API, sold, or sub-licensed.
Bu veriyle bir model eğitirseniz, modeli kamuya açık şekilde yayınlamayı (açık ağırlık veya herkese açık demo) ve bu veri kümesine atıf vermeyi kabul edersiniz. / If you train a model on this data, you agree to publish it publicly (open weights or a public demo) and to cite this dataset.
Sesleri ve transcript'leri yeniden dağıtamazsınız; erişim kişiseldir, ekibiniz dışına aktarılamaz. / Do not redistribute the audio or transcripts; access is personal and non-transferable.
Belirli bir kişinin sesini taklit etmek, kimliğini tespit etmek veya rızası olmadan ses klonu üretmek için kullanamazsınız. / Do not use the data to impersonate or identify a specific person, or to clone anyone's voice without their consent.
Kaynak kayıtların hakları temizlenmemiştir; kendi hukuki sorumluluğunuzu kabul edersiniz ve hak sahibi talebi üzerine kopyalarınızı silersiniz. / The rights of the source recordings are not cleared; you accept sole legal responsibility for your use and will delete your copies upon a rightsholder request.

Formdaki beyanların doğru olmadığı anlaşılırsa erişim iptal edilir. / Access is revoked if the form turns out to be inaccurate.

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.

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 a single flag: suspected synthetic speech, from a check that is deliberately paranoid and also fires on clean studio narration. 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. Transcript'ler whisper-large-v3-turbo ile üretildi ve ikinci bir geçişle karşılaştırıldı; insan doğrulaması yapılmadı. 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 tek bir sebeple, "sentetik ses şüphesi" ile işaretli ve bu kontrol temiz stüdyo anlatımını da yakalıyor. Kontrol etmeden eğitime katmayın.

Kaynak kayıtların yayın hakları doğrulanmamıştır; erişim onaya tabidir ve ticari kullanıma kapalıdır (bkz. Access and terms).

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 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.

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)

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 69% of clips fall in the 9–13 s band.

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

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	2.5–20.0 s	20.0–21.0 s boundary cases	< 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
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

Because every metric is stored per row, you can reproduce a stricter or looser cut yourself 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. Expect a large share of these to be real human readings. That is a hypothesis worth testing, not a promise; some of them really are TTS.

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) 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"]))

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.

Türkçe

review, otomatik filtrelerin emin olamadığı 432.306 klip (1.412 saat). Çöp değildir; kesin elenenler (21.926 klip) zaten bu veri kümesinde yok. Havuzun %83'ü tek bir sebeple işaretli: AudioSet'in "sentetik ses" etiketi, ki bu etiket temiz stüdyo anlatımında da tetikleniyor. 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 to match the VoxCPM2 AudioVAE training input; resample if your model expects 22.05/24/44.1 kHz.

Limitations and caveats

No human verification anywhere. Transcripts are machine-generated. 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 unverified by construction and must not be treated 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 (see Splits).
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

This is a gated dataset: fill in the access form on this page and the request is reviewed manually. Access is granted for non-commercial research and open model development, on these terms:

No commercial use — not the data, not derivatives, not models trained on it: no paid product, service, API, resale or sub-licensing.
Publish what you build — if you train a model on this data, release it publicly (open weights or a public demo) and cite this dataset.
No redistribution — access is personal and non-transferable; do not re-upload the audio or transcripts.
No voice impersonation — do not clone or identify a specific person's voice without their consent.
Own your legal risk — see Licensing; delete your copies on a rightsholder request.

Access is revoked if the information in the form turns out to be inaccurate.

Erişim onaya tabidir. Şartlar: ticari kullanım yok (veri, türevleri ve bununla eğitilen modeller ücretli ürün/hizmet/API'de kullanılamaz, satılamaz), bu veriyle model eğitirseniz modeli kamuya açık yayınlama ve bu veri kümesine atıf verme, veriyi yeniden dağıtmama, belirli bir kişinin sesini rızası olmadan klonlamama, ve hukuki sorumluluğu üstlenip hak sahibi talebinde kopyaları silme.

Licensing

license: other. The source recordings were collected from publicly accessible Turkish audiobook and spoken-article uploads; their individual copyright status was not cleared, and automatic processing grants no redistribution rights. Whoever uses or publishes this dataset — 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 — pipeline voxcpm/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: 394

Total file size:

172 GB

Erişim onay gerektirir / Access requires approval