No dataset licence is declared. The audio originates from public YouTube uploads, whose public availability does not establish redistribution or commercial-use permission. Credit the source channels and the dataset. Recording, literary-text and narrator rights must be assessed separately; this card does not grant rights held by third parties. Rights holders can request removal through repository issues or dataset discussions.

KIRAAT — A Turkish Read-Speech Corpus

A sentence-aligned read-speech corpus built from publicly available recordings on Turkish audiobook YouTube channels. The channel credits are in the table at the end of this card; every clip carries the channel it came from in the channel column.


clips	1,840,404
duration	3,105.7 hours
recommended subset	1,547,494 clips / 2,575.2 hours
channels	27
speaker clusters (inferred)	90
source recordings	2,680
words (ASR)	21,695,774
audio	24 kHz, mono, FLAC, embedded in the parquet shards
download size	~250 GB (clip files measured at 246 GB)
splits	`train` / `dev` / `test` / `rest` — the whole corpus is published
language	Turkish

Loading

from datasets import load_dataset

# the default training subset
train = load_dataset("serdarcaglar/kiraat", split="train")

# 250 GB is a lot to download for a look; stream instead
train = load_dataset("serdarcaglar/kiraat", split="train", streaming=True)
clip = next(iter(train))
clip["audio"]["array"], clip["text_spoken"], clip["speaker_id"]

# everything the policy left out — published, not deleted
rest = load_dataset("serdarcaglar/kiraat", split="rest", streaming=True)

Audio is embedded in the parquet shards, so there is no separate audio download; datasets decodes the audio column for you. This needs datasets>=4 — the shards were written with the List feature type, and on 3.x loading fails with Feature type 'List' not found before any row is read. Decoding the audio column also needs torchcodec installed; without it you can still read the raw FLAC bytes with .cast_column("audio", Audio(decode=False)).

Two things that set this corpus apart

Segmentation follows transcript sentence boundaries. The reported 206-recording comparison uses lowercase starts and missing final punctuation as transcript proxies. Lowercase-start rates are 14.31% in the implemented VAD baseline and 1.39% in the sentence arm; unflagged rates are 12.81% and 0.00%. These checks depend on ASR casing/punctuation and do not establish acoustic truncation rates. Baseline clips can be inspected from their text, so absence of a dedicated flag does not make the proxy undetectable.

No threshold ever drops a clip. Quality measurements are published as columns. Which clips belong to the default training subset is decided by a versioned policy, and the reason for every exclusion is written into the exclusion_reasons column. Small score-stratified listening audits informed removal of two candidate thresholds (clip_ratio, word_confidence). They do not establish corpus-wide error rates or downstream training utility. You can cut your own thresholds from the columns — clips with recommended=false have not been deleted.

Splits

split	clips	hours	recordings	channels	speaker clusters
train	1,512,448	2,517.08	2,579	27	89
dev	5,750	9.54	40	27	26
test	5,473	9.30	45	27	26
rest	316,733	569.78	—	27	—
total	1,840,404	3,105.70	2,680	27	90

train, dev and test are drawn from the recommended subset. Every remaining clip is published in a fourth split, rest — nothing is deleted, and a clip's measurement columns are the same there as anywhere else:

what is in `rest`	clips	hours
`recommended=false`	292,910	530.5
recommended but unused: beyond a channel's dev/test duration target, or dropped by the text-leakage cleanup (these are never moved into `train`)	23,823	39.3

The recordings column above counts recordings whose clips form a split; rest draws clips from recordings across all three, and 16 of the 2,680 recordings have all of their clips in rest.

The train/dev/test splits have disjoint source IDs. rest also contains clips from their recordings. If adding rest to training, first exclude every dev/test source ID and recheck transcript overlap. dev and test are channel-balanced (equal duration target per channel). Text leakage was checked and cleaned: dev/test clips carrying a sentence identical to one in train were dropped, and the remaining overlap is zero for train--dev and train--test. A separate database audit found 12 shared normalized text keys between dev and test. These checks do not exclude near-duplicate text, shared books or re-uploads under different IDs.

The reported cluster assignments have overlap: 25 of the 26 test clusters also occur in training. These are inferred speaker labels, not manually verified identities, and the evaluation is predominantly seen-cluster. If you want an unseen-speaker experiment, build your own split from the speaker_id column.

How it was built

Decode to 24 kHz mono with ffmpeg → long-form ASR with faster-whisper large-v3 (word timestamps) → forced alignment with torchaudio MMS_FA → channel-level boilerplate/announcement mining → sentence-boundary segmentation with audio-based boundary refinement → clip measurements (VAD speech ratio, silences, level, BS.1770 loudness; music via HDemucs + AudioSet; DNSMOS P.835; speaker embeddings via ECAPA-TDNN) → recommended subset via the versioned policy.

The export code writes model identifiers, configuration and stage versions to run.json. Speaker-weight provenance is incomplete: the configured revision is not passed to the loader, and the provenance record omits the speaker model and its dependencies. The pipeline source is available: https://github.com/serdarildercaglar/kiraat. Every count, hour and share printed in this card is regenerated from the run database by scripts/verify_card.py — including the duplicate marking, the policy decision and the four splits, which are recomputed rather than read back.

Quality distributions (recommended subset)

column	p05	p50	p95
duration (s)	2.76	5.82	10.56
`loudness_lufs`	−27.99	−20.92	−12.54
`dnsmos_ovrl`	2.862	3.296	3.491
`dnsmos_sig`	3.271	3.563	3.710
`dnsmos_bak`	3.631	4.109	4.210
`speech_ratio`	0.851	0.977	1.000
`align_score_mean`	0.858	0.961	0.994
`word_confidence`	0.550	0.917	0.999
`n_words`	5	11	21

Across the whole corpus, 83.6% of clips have dnsmos_ovrl ≥ 3.0 — a threshold commonly used in the field (at ≥ 3.5 the share is 3.9%). No loudness normalization is applied. Mono conversion, resampling and a 40 Hz high-pass filter do modify the signal. Level remains a column.

Music separation was skipped for 1,455,201 clips and run on 385,203 clips. music_to_speech_db = -80 with music_db_separated=false is an unmeasured sentinel, not evidence of absent music. Policies below the screening threshold need further inference.

Recommended-subset policy (v6)

For recommended=true: speech_ratio ≥ 0.60, internal_silence_sec ≤ 1.0, and none of the flags oversize, forced_split, gap_split, short, duplicate, boilerplate, background_music.

This leaves out 292,910 clips (530.5 hours). What each rule contributes to that total — a clip can fail more than one rule:

rule	clips excluded
flag present	281,995
`internal_silence_sec > 1.0`	12,757
`speech_ratio < 0.60`	2,412

Flag frequencies over the whole corpus: background_music 207,196 (11.26%), forced_split 52,501 (2.85%), short 20,904 (1.14%), oversize 6,731 (0.37%), gap_split 5,141 (0.28%), boilerplate 1,053 (0.06%), duplicate 576 (0.03%).

The duplicate heuristic matches normalized text, duration, LUFS, RMS and peak exactly. Matching summary values do not prove waveform identity, and re-encoding or shifted boundaries can evade the heuristic. Its precision and recall have not been validated; separate readings are intended to be retained.

Columns

column	type	unit	stage	description
`audio`	audio	24 kHz	export	The clip audio itself, mono FLAC embedded in the parquet shards. `datasets` decodes it to `{array, sampling_rate}`; there is no separate audio download. No gain is applied — see `loudness_lufs`.
`id`	string		segment	Clip id, `srcNNNNN-KKKKK`: the source recording's number and the clip's index within that recording.
`source_id`	string		prepare	The source recording the clip was cut from (`srcNNNNN`). Every clip of one recording carries the same value; this is the key to group by if you build your own recording-level splits.
`channel`	string		prepare	The YouTube channel the source recording came from (folder name). This is not a speaker identity; a channel may have several readers.
`source_sample_rate`	int	Hz	prepare	The source recording's actual container sample rate. Preparation resamples to at most 24 kHz without upsampling lower-rate sources. All 2,698 prepared recordings in the stored run report 24 kHz; dataset loading declares 24 kHz audio.
`source_flags`	list[string]		prepare	Recording-level flags. `truncated_source`: the decoded audio is shorter than 95% of the duration the container reports (a corrupt or partial download).
`start`	float	s	segment	Clip start time within the source recording, taken from the word timestamp at the sentence boundary, with boundary refinement and lead padding applied.
`end`	float	s	segment	Clip end time within the source recording, same as `start`.
`duration`	float	s	segment	Clip duration, `end - start`.
`text_raw`	string		asr	Untouched ASR output (Whisper large-v3), joined over the clip's word span.
`text`	string		segment	Lightly cleaned text: whitespace, quotes and repeated punctuation fixed; numbers and abbreviations left as they are. The readable/archival version.
`text_spoken`	string		segment	Text converted to how it is spoken: numbers, ordinals, times/dates/currencies and common abbreviations written out in Turkish. This is the field intended for TTS training.
`n_words`	int		segment	Word count in the clip (ASR words, with clitics — de/da, mi, ki — attached to the preceding word).
`word_confidence`	float		segment	The lowest word probability in the clip, either a Whisper word probability or an aligner score depending on `confidence_source`. A single rare word (proper noun, interjection) drags it down; it is a ranking signal, not a gate.
`word_confidence_mean`	float		segment	The clip mean of word probabilities from the same source.
`confidence_source`	string		segment	Where the `word_confidence` columns come from: `asr` (Whisper word probability) or `align` (forced-alignment score).
`align_score_min` (optional)	float		align	The minimum MMS_FA CTC word alignment score in the clip; a model score, not a calibrated probability of transcript correctness or a human-reference error measure.
`align_score_mean` (optional)	float		align	The clip mean of the aligner's word scores.
`lead_gap_sec` (optional)	float	s	segment	Gap between the clip's first word and the previous word in the recording. Empty if it is the recording's first word. A small value means the previous sentence risks bleeding into the clip.
`trail_gap_sec` (optional)	float	s	segment	Gap between the clip's last word and the next word in the recording; empty if it is the recording's last word.
`leading_silence_sec`	float	s	clip_qc	Time from the clip start to the first speech found by VAD (Silero), from a separate no-padding, short-silence-threshold VAD pass.
`trailing_silence_sec`	float	s	clip_qc	Time from VAD's last speech to the clip end, from the same no-padding pass.
`internal_silence_sec`	float	s	clip_qc	The longest speechless stretch inside the clip (largest gap between VAD regions).
`speech_ratio`	float		clip_qc	The fraction of clip duration VAD counts as speech (0–1).
`peak_dbfs`	float	dBFS	clip_qc	Absolute peak sample level.
`rms_dbfs`	float	dBFS	clip_qc	RMS level of the clip (no gain applied — the source's own level).
`clip_ratio`	float		clip_qc	Fraction of samples at full scale (\|x\| ≥ 0.99); a measure of digital clipping.
`loudness_lufs` (optional)	float	LUFS	clip_qc	BS.1770 integrated loudness (pyloudnorm). No gain is applied to the audio; level normalization is left to you, and this column is its input. Empty for clips shorter than the 0.4 s measurement block or entirely silent.
`music_score_audioset` (optional)	float		music	Maximum score of an AudioSet AST classifier over its music labels (Music, Background music, Soundtrack…), taken as the max over clip windows. A cheap screening signal; it can coexist with speech. Empty for unreadable clips (`unreadable_audio`).
`music_to_speech_db` (optional)	float	dB	music	A physical measurement from source separation (HDemucs): the ratio of accompaniment energy (drums+bass+other) to vocal energy. If the separator did not run, the floor value −80. The `background_music` flag requires both the dB threshold and the configured minimum AudioSet score. The floor value does not measure absent music. Empty for unreadable clips.
`music_db_separated` (optional)	bool		music	Whether the separator actually ran on this clip. If not, `music_to_speech_db` is the floor value rather than a measurement (the AudioSet score was below the screening threshold). Empty for unreadable clips.
`music_stem_db` (optional)	string (JSON)	dB	music	Per-stem energy levels for the separator's four stems (drums, bass, other, vocals); written only for clips the separator ran on. `music_to_speech_db` is derived from these. Published as a JSON object encoded in a string (the key set is not fixed), so parse it with `json.loads`.
`dnsmos_sig` (optional)	float		dnsmos	DNSMOS P.835 speech quality estimate (SIG, 1–5): distortion of the speech itself. Matches the reference implementation exactly: 9.01 s windows, polynomial mapping, window averaging. Empty for unreadable clips (`unreadable_audio`).
`dnsmos_bak` (optional)	float		dnsmos	DNSMOS P.835 background estimate (BAK, 1–5): how intrusive the background noise is; higher means cleaner. Empty for unreadable clips.
`dnsmos_ovrl` (optional)	float		dnsmos	DNSMOS P.835 overall quality estimate (OVRL, 1–5). Not a gate: a policy rule may only be added after blind listening review (see the policy v4 record). Empty for unreadable clips.
`flags`	list[string]		export	Clip flags; none of them removes a clip. `forced_split`: a sentence that exceeded the duration ceiling on its own was split at internal punctuation; `gap_split`: the same sentence was split at a long silence; `oversize`: no internal punctuation to split on, so the sentence was kept whole and exceeds the ceiling; `short`: duration below `segment.min_sec`; `background_music`: both music thresholds pass; `duplicate`: normalized text and audio summary values match (see `duplicate_of`); `boilerplate`: an intro/announcement pattern that recurs across the channel's recordings; `unreadable_audio`: the clip file could not be read and has no measurements.
`duplicate_of` (optional)	string		export	If the clip is a duplicate, the id of the copy that was kept; otherwise empty. The heuristic matches normalized text and `dedupe.identity_fields` (duration, LUFS, RMS, peak) exactly; this does not prove waveform identity. A separate reading of the same text is NOT a duplicate — it is kept as prosodic variety; channel is not part of the key.
`speaker_id` (optional)	string		export	The recording's speaker cluster (`spkNNNN`). Clustering runs over the whole corpus: the merge threshold is taken from the equal-error point of two measured distributions (within-recording clip similarity, and cross-channel recording similarity), and no cluster count is imposed. The identity is recording-level; all clips of a recording carry the same value. Not a gate — the policy has no rule on it.
`speaker_consistency` (optional)	float		speaker	Mean pairwise cosine similarity among the recording's embedded clips. A low value may indicate acoustic or speaker variation; it does not by itself prove multiple speakers or substitute for diarization.
`speaker_margin` (optional)	float		export	The recording's similarity to its own cluster centroid minus its similarity to the nearest other cluster. A small value points to a recording near a cluster boundary; a negative value says it may be in the wrong cluster.
`recommended`	bool		export	Whether the versioned policy (`recommended_subset`) recommends this clip for the default training subset. It removes no data; you can set your own rule from the measurement columns.
`exclusion_reasons`	list[string]		export	If `recommended=false`, the list of rules that were not met: `metric<threshold`, `metric>threshold`, `flag:name[,name]`, `missing_measurement:metric`. Empty for recommended clips.
`policy_version`	string		export	The policy version that produced `recommended` and `exclusion_reasons`; the rules are versioned in the config, and when they change only these three columns are recomputed.
`split`	string		export	The published split the row belongs to: `train`, `dev`, `test` or `rest`.

Known limitations

The text is ASR output, not corrected by hand; text_raw, text and text_spoken are all derived from the same ASR transcript.
Digits are not force-aligned. The aligner (MMS_FA) is character-based and its dictionary has no digit characters, so words containing digits ("1923", "15'inci", "%20") get their timestamps from ASR rather than the aligner, and those words do not contribute to align_score. Share of affected clips: 3.08% (2.91% within the recommended subset). Converting the text to its spoken form before alignment would have solved this, but it would change the published text itself, so it was not done; the text_spoken column provides the spoken form separately.
Garbage alignment on short sentence-initial words occurs (measured share 2.12%); align_score_min makes it visible.
Channel and speaker concentration: in the recommended subset the top two channels account for 38.2% of the hours, and the top five of the 90 speakers carry 54.4% of the hours.
A channel is not a speaker: 13 of the 27 channels have more than one reader. Speaker identity comes from clustering and is recording-level.
A single encoding profile: sources are ~129 kb/s AAC; the median source bandwidth cut is 15.7 kHz, and around 13 kHz on two channels.
The test set has not been verified by hand; the leakage check is automatic.
2,699 recordings were downloaded and 2,680 produced clips: one could not be decoded, and 18 more yielded no sentence-bounded clip.

Channel credits

Built from the publicly available recordings of the channels below. Work using this corpus is expected to carry these credits.

channel	recordings	clips	hours	recommended hours	speaker clusters
seslikitaplarmavi	279	332,847	580.7	539.5	2
BirDinle	235	285,797	487.9	444.8	6
dinleyiniz	83	138,254	239.4	223.2	32
sess-Seslikitap	176	149,786	226.2	199.2	4
Peri_Mia	216	132,779	206.9	80.1	1
Pandoramedyaseslikitap	67	112,170	193.6	91.0	1
ZubeyirSener	99	110,294	185.6	148.5	1
ses-arşiv	234	92,190	154.6	142.6	7
cantadakitap	291	74,328	126.4	111.9	1
anahtarca	175	61,242	111.7	92.5	3
kitaplar	146	63,096	103.0	73.4	2
seslikutuphanemkanali	50	45,723	77.9	74.4	1
kitapdinle	33	41,540	72.0	68.6	1
idea_stüdyo	43	33,102	54.9	52.6	1
MuratKaraOfficial2021	71	24,059	39.3	34.9	1
seslimakalem	182	19,144	37.5	30.9	1
SesliKitaPodcast	41	18,232	31.1	22.7	1
seslikitapturkish	27	17,238	28.9	25.8	7
sesli-kitaplar	44	13,255	24.5	15.3	3
eba	39	15,104	24.4	23.4	9
Seslendiriyor	28	14,262	24.1	20.2	7
OkumaSaati	58	14,889	22.0	21.5	1
SESLİKİTAPEVİ	29	13,041	21.2	8.5	1
denizinötesindekisesler	16	7,786	14.8	13.4	1
seskitap	12	5,558	9.1	8.7	2
bizimkütüphane	3	4,024	6.9	6.7	1
KitaplarinKedisi	3	664	1.0	1.0	2
total	2,680	1,840,404	3,105.7	2,575.2	90

The speaker column counts the distinct clusters heard on each channel, so it adds up to more than 90: a cluster can appear on several channels — the same reader published by more than one of them. The hours columns are rounded per channel and add up to 3,105.6 / 2,575.3.

Rights, use and removal

No dataset license is declared. Public availability of the source uploads alone does not establish permission to redistribute their audio or literary texts. This card does not grant rights held by recording, text or voice rights holders, and does not establish permission for commercial use or voice cloning. Licenses for code, derived annotations and underlying recordings must be specified separately once the applicable rights have been documented.

Credit the dataset and the source channels in the table above. Attribution and an access acknowledgment do not substitute for a redistribution license or narrator consent. The Hugging Face access conditions apply separately.

Rights holders and narrators can request removal through an issue in the code repository or a discussion on the dataset page. The stated procedure is to remove the affected channel's recordings; the release should additionally record affected clip IDs and dataset revisions. Source-by-source permission records and versioned removal records are not supplied in this repository.

Personal and sensitive information

The audio is human speech, so it carries identifiable voices: 90 speaker clusters read these recordings. speaker_id is a cluster derived from speaker embeddings, not a name — no identity, contact or demographic information is published, and no attempt was made to link a cluster to a person. The material is read literature from public uploads rather than private conversation, but that does not rule out personal or sensitive information in the text; the transcripts were not screened for it by hand. Requests to remove a voice are handled the same way as rights holder requests above.

Citation

Serdar İlder Çağlar — ORCID 0000-0002-5776-2431

@misc{kiraat2026,
  title  = {KIRAAT: A Sentence-Aligned Turkish Read-Speech Corpus},
  author = {Serdar {\.I}lder {\c{C}}a{\u{g}}lar},
  orcid  = {0000-0002-5776-2431},
  year   = {2026},
  url    = {https://github.com/serdarildercaglar/kiraat},
  note   = {ORCID: https://orcid.org/0000-0002-5776-2431}
}

Downloads last month: 963

Total file size:

293 GB

Models trained or fine-tuned on serdarcaglar/kiraat