gaydmi/emo_tts

Tatar Dubbed Speech

Sentence-level speech segments in Tatar, cut from Tatar-language dubs and aligned to their subtitles with CTC forced alignment.

7,257 sentences / 4:39:06 drawn from 11.85 h of source audio — dialogue is sparse in this material, so roughly 40% of the runtime is speech.

Source	source prefix	Rows	Duration	Median similarity
Берсерк, 25 episodes	Берсерк - N серия	5,855	4:03:34	0.941
Мистер һәм миссис Смит	Мистер һәм миссис Смит	1,402	0:35:32	0.909

Loading

from datasets import load_dataset

ds = load_dataset("gaydmi/emo_tts", split="train")

# segments whose audio was verified to contain the transcript
clean = ds.filter(lambda r: r["similarity"] >= 0.8)

Data fields

• audio — the segment, mono 16-bit PCM WAV @ 32 kHz
• text — one sentence
• duration — seconds
• file_id — source audio file (episode or film)
• id — zero-padded index of the segment within its file, in reading order
• source — title, per the table above
• speaker_name — character name from the subtitles where present, else "" (48% of rows)
• speaker_cluster — model-derived speaker grouping, see below
• similarity, score, drift, lead_silence, trail_silence — see Quality signals

Segmentation

Rows are sentences, not raw subtitle cues. Consecutive cues are joined into runs — broken by a gap of more than 2 s, or by a speaker change where the subtitles record one — and each run is re-split with razdel. So a sentence spread across three cues becomes a single row, and a cue holding two sentences becomes two rows.

The two titles move in opposite directions under this, which is the intended behaviour: the anime subtitles spread one sentence over several short cues (6,299 cues → 5,855 sentences), while the film packs several sentences into one cue (1,144 → 1,402).

A margin of up to 200 ms is kept either side of each cut, capped at half the gap to the neighbouring sentence so clips never overlap. Median lead_silence is 140 ms and trail_silence 120 ms. Sentences are short: median 1.76 s and 22 characters, 44% at or above 2 s.

Quality signals

Nothing is filtered. Every aligned sentence is present, and the measurements ship as columns so you can pick your own cut.

Column	Meaning	Median	p90
similarity	character overlap between text and a greedy CTC decode of the cut audio	0.938	1.000
score	mean CTC log-probability of the aligned path	-0.09	-0.01
drift	aligned start minus the subtitle's own start, seconds	0.07	0.96
lead_silence / trail_silence	silence at each edge, seconds	0.14 / 0.12	0.21 / 0.19

similarity is the one to filter on — it asks directly whether the audio contains the words claimed for it, so it catches segments that aligned confidently but wrongly.

Filter	Rows	Duration
none	7,257	4:39:06
similarity >= 0.7	6,745	4:28:03
similarity >= 0.8	6,310	4:15:24
similarity >= 0.9	4,908	3:25:33

drift is a weak signal on its own — segments that moved 1–3 s are no worse than ones that did not — but anything past 10 s is unusable.

speaker_cluster

Speaker embeddings from speechbrain/spkrec-ecapa-voxceleb (ECAPA-TDNN), clustered agglomeratively per title, giving 200 groups. Берсерк is clustered across all 25 episodes at once, so a recurring character keeps one id throughout; the film is clustered separately. Clips of 2 s or more set the clusters; shorter ones are attached to the nearest centroid.

These are derived labels, not verified identities. Measured against the character names the subtitles do carry (used only to choose and check the threshold, never copied into the output):

	V-measure	homogeneity	completeness
Берсерк, clips ≥2 s	0.811	0.823	0.799
Берсерк, shorter clips	0.659	0.789	—
Берсерк, whole column	0.703	—	—
Film, clips ≥2 s	0.738	0.719	0.758
Film, whole column	0.651	—	—

Speaker identity is recoverable on longer clips and weak below 2 s. At homogeneity around 0.7 a given id still mixes voices, so the column suits filtering and analysis rather than speaker-conditioned synthesis. This is a limit of short clips with a musical score underneath, not something further tuning resolves. Gate on duration >= 2 for the more reliable subset.

Processing

• Subtitles parsed from .srt and .ass; override tags, HTML and music glyphs stripped.
• Cues joined into runs and re-split into sentences with razdel.
• Each file aligned against those sentences with ctc-forced-aligner and the default MahmoudAshraf/mms-300m-1130-forced-aligner model, romanising via uroman (tat).
• star_frequency="segment", so the aligner can skip music, effects and untranscribed audio.
• Aligned in ~5-minute chunks cut at subtitle gaps, keeping the forced-align trellis small.
• Audio cut at the aligned boundaries plus margins, resampled to 32 kHz mono.

Usage considerations

• This is dubbed media, not studio narration. Music and sound effects sit under much of the dialogue. No column here detects that — a line over a musical score can score similarity 1.00. For TTS this is likely the binding quality constraint rather than transcript accuracy, and is worth a listening pass before trusting the hours.
• Sentences are short — median 1.76 s. Useful for ASR, thin for prosody modelling.
• speaker_name comes from the subtitles, covers 48% of rows across 101 names, and has not been verified against the audio.
• The film yields the least and scores lowest (0:35:32 from a 2-hour runtime, median similarity 0.909 against 0.941 for the series); it sits in its own folder so it is easy to exclude.

Source material and rights

The audio is derived from Tatar-language dubs of two commercial works, Berserk (1997 TV series) and Mr. & Mrs. Smith (2005). The rights to the underlying works are held by their respective owners, and no license is asserted over them here. This dataset is published for Tatar language-technology research; if you are a rights holder and want it taken down, open a discussion.