Crimean Tatar Audiobooks

Dataset Summary

Crimean Tatar Audiobooks is a speech dataset sourced from different sources (public radio stations/youtube channels) containing audiobooks in Crimean Tatar. The dataset comprises recordings of different native fiction books, all read by a single female speaker (for now). The dataset is intended for text-to-speech (TTS) research and development in the Crimean Tatar language.

25:34:53 of aligned speech in 22,517 segments, across 9 book parts by 3 authors.

Dataset Structure

Parts: The dataset contains one configuration:

• audiobooks: the audiobook recordings, segmented into sentences.

Each book part lives in its own folder, sharded as train_{i:05d}_of_{n:05d}.parquet (max 1500 rows per shard).

Book	source	Segments	Duration
Черкез-Али — Сабалар къучагъында, 1	Черкез-Али - Сабалар къучагъында. Биринджи къысым.	3,295	3:31:49
Черкез-Али — Сабалар къучагъында, 2	Черкез-Али - Сабалар къучагъында. Экинджи къысым.	3,677	4:10:20
Черкез-Али — Сабалар къучагъында, 3	Черкез-Али - Сабалар къучагъында. Учюнджи къысым.	3,056	3:22:32
Черкез-Али — Сабалар къучагъында, 4	Черкез-Али - Сабалар къучагъында. Дёртюнджи къысым.	1,664	2:02:34
Шамиль Алядин — Теселли	Шамиль Алядин - Теселли	2,707	2:54:29
Шамиль Алядин — Чауш огълу	Шамиль Алядин - Чауш огълу	2,396	2:20:28
Урие Эдемова — Баш язысы, 1	Урие Эдемова - Баш язысы. 1 къысым	2,682	3:32:04
Урие Эдемова — Баш язысы, 2	Урие Эдемова - Баш язысы. 2 къысым	1,442	1:53:06
Урие Эдемова — Баш язысы, 3	Урие Эдемова - Баш язысы. 3 къысым	1,598	1:47:30
Total		22,517	25:34:53

Data Fields:

• audio: the audio segment (mono WAV, 16-bit PCM, 32 kHz).
• text: the corresponding transcription.
• duration: the duration of the audio segment in seconds.
• id: a unique identifier for each segment within a file (zero-padded, in reading order).
• file_id: the identifier of the source audio file (chapter) within the book part.
• source: the author and title of the book.
• speaker_id: the id of the reader.
• similarity: per-segment alignment confidence, 0-1. See Quality signals below.

Data Processing

Text Processing:

• Book texts were converted to Markdown and split per chapter to match the audio files.
• Markdown/HTML leftovers, footnote references and backslash-escaped punctuation were stripped.
• Text was split into sentences using the razdel library.

Audio Processing:

• Each sentence was aligned to its chapter audio with the ctc-forced-aligner library using the default MahmoudAshraf/mms-300m-1130-forced-aligner model, romanizing the text with uroman (crh).
• Audio was cut at the resulting sentence boundaries and resampled to 32 kHz mono.

Technical Details:

Dataset Type	speech corpus for TTS
Language	crh, Crimean Tatar
Speech Style	scripted monologue
Content	fiction books
Audio Parameters	32 kHz, 16-bit, mono
File Format	WAV (PCM) in Parquet
Recording Environment	quiet indoor environment

Total Duration:

• Books: 25:34:53 (25 hours, 34 minutes, 53 seconds)

Quality signals

Every segment carries a similarity score: the character overlap between its text and a greedy CTC decode of its own audio, using MahmoudAshraf97/mms-300m-1130-forced-aligner. It answers a direct question -- does this clip actually contain the words we claim are in it? -- so it catches segments that alignment placed confidently but wrongly. Nothing has been filtered on it; the column is there so you can choose your own cut.

Overall median is 0.912, but it varies a lot by book:

Book	Segments	Median	% below 0.5	% below 0.8
Шамиль Алядин - Теселли	2,707	0.848	13.2%	37.8%
Урие Эдемова - Баш язысы. 3 къысым	1,598	0.891	21.0%	31.4%
Шамиль Алядин - Чауш огълу	2,396	0.901	2.4%	16.5%
Черкез-Али - Сабалар къучагъында. Учюнджи къысым.	3,056	0.914	10.2%	18.0%
Черкез-Али - Сабалар къучагъында. Дёртюнджи къысым.	1,664	0.918	0.5%	9.4%
Урие Эдемова - Баш язысы. 1 къысым	2,682	0.920	3.1%	11.4%
Урие Эдемова - Баш язысы. 2 къысым	1,442	0.921	2.1%	9.4%
Черкез-Али - Сабалар къучагъында. Экинджи къысым.	3,677	0.923	0.4%	8.1%
Черкез-Али - Сабалар къучагъында. Биринджи къысым.	3,295	0.925	0.7%	7.6%
All	22,517	0.912	5.4%	16.1%

Keep-rates for a global threshold:

Filter	Rows	Hours
none	22,517	25:34:53
similarity >= 0.7	20,491	23:57:26
similarity >= 0.8	18,901	22:36:02
similarity >= 0.9	12,795	16:11:08

Because the spread between books is wider than the effect of any single cutoff, a global threshold over-prunes the strong books in order to clean up the weak two. Filtering per book, or simply excluding the weakest, will usually retain more usable audio.

Two rows have a null similarity, where the text normalised to nothing.

There is deliberately no CTC alignment-score column. Computing one requires ctc_forced_aligner's forced_align_cpp, which writes out of bounds on some inputs in this data and corrupts the process heap -- it crashes at unrelated call sites with SIGSEGV, SIGBUS, SIGTRAP or SIGABRT, and segfaults outright whenever a clip has fewer frames than its text has tokens. The similarity score above never calls it.

Usage Considerations

• The dataset was not extensively preprocessed. Users are encouraged to perform additional preprocessing (e.g. normalization, cleaning, or re-segmentation) as needed for their specific applications.
• No alignment-quality filtering has been applied. Segment durations range from under 0.1 s to 65 s (median 3.4 s). Where the book text and the recording diverge, forced alignment stretches or compresses segments to absorb the difference, so some rows do not match their audio. Filtering on implausible characters-per-second (len(text) / duration, normally ~13–15) removes most of these.
• All texts are written in Cyrillic script.
• Known problem areas:
  • Шамиль Алядин - Теселли is the weakest book by a wide margin: median similarity 0.848, with 13.2% of its segments below 0.5 against 0.4-0.7% for the strongest books. Filter it hard or skip it.
  • In Урие Эдемова - Баш язысы. 3 къысым, file_id 38 is largely misaligned (its chapter text contains considerably more material than the recording) and is best dropped. file_id 34 has one bad stretch, as does file_id 20 in part 1 (its chapter text runs past the end of the recording).

Citation

If you use this dataset in your work, please, cite us.

License

This dataset is distributed under the CC BY-SA 4.0 license.