tamil-turns

4,774 conversational turns from 115 real Tamil telephone calls — one speaker's continuous hold of the floor, with their own mid-turn pauses kept separate from the silence that ends the turn.

Companion to santhosh-005/tamil-eot, which is the same 115 calls cut as 8 s windows for binary end-of-turn classification. This one keeps whole turns, so it carries the structure an 8 s window destroys by construction: where the speaker paused mid-thought, where the other speaker overlapped them, and how long the floor stayed empty.

from datasets import load_dataset
ds = load_dataset("santhosh-005/tamil-turns", split="test")

What it is good for

end-of-turn / endpointing	the last silence_spans entry is the real end-of-turn
mid-turn pauses	every earlier span is the speaker hesitating, not finishing
overlap and interruption	speech_crosstalk measures the other leg talking across the turn; a negative eot_gap means they started before this speaker stopped
ASR / timing	speech_s, word_rate, n_spans and the span list give per-turn timing

Schema

field	meaning
id	<call>_<L|R>_<turn_index>
audio	16 kHz mono FLAC — that speaker's leg only, never a mixdown
language	ta
silence_spans	{start, end} in clip time. The last is the end-of-turn; earlier ones are mid-turn pauses
messages	the other speaker's preceding turn
duration	speech portion, before the end-of-turn silence
eot_gap	true length of the final silence. Negative would mean overlap — excluded here by the quality gate
speech_s	seconds actually spoken (duration minus the pauses)
word_rate	n_words / speech_s
n_spans, n_absorbed	VAD spans in the turn; backchannels folded in
pause_crosstalk, speech_crosstalk	other-leg speech inside a pause / across the whole turn
transcript	words apportioned to the turn — approximate, see below
conversation_id, speaker_id, turn_index	conversation structure — sort by turn_index to rebuild a call
split	train / dev / test, by call
eot_bench_eligible	meets eot-bench's release rules (4,322 rows)
label, source, sid	joined from tamil-eot where a labelled boundary lands on this turn's end; null otherwise

split	turns	eligible
train	3,024	2,739
dev	482	437
test	1,268	1,146

Split by call — no call appears in two splits, and the split is the same one tamil-eot uses, so the two datasets agree on which 30 calls are held out. 2,113 turn ends carry a label joined from tamil-eot.

Use with livekit/eot-bench

silence_spans follows the eot-bench convention — last span is the end-of-turn, earlier ones are holds — so this is a drop-in subset:

eot-harness predict --path santhosh-005/tamil-turns --name default --split test \
  --adapter eot_harness.smart_turn_adapter:SmartTurnAudioAdapter --output-dir out

Filter to eot_bench_eligible to match their release rules exactly: final silence ≥ 0.2 s so every turn is scorable at the 0.2 s score point, and no span

5.0 s. The 452 excluded rows are fast handovers and long lapses — deliberately kept in the dataset because interruption and timing work wants them.

Tamil is not in eot-bench's BENCHMARK_LANGUAGES, so ta must be added there or every adapter reports unsupported_language and silently scores nothing.

How turns were built

Turns come from floor occupancy over VAD spans confirmed against the human transcript, never from a turn classifier. Maximal same-side runs, with backchannels absorbed (< 3 words and < 1.0 s) and a split at any pause that is not a hesitation — over 2.0 s with the floor empty, or one the other speaker talks across.

Six gates then decide clean, and only clean turns are published: the turn opens and closes a transcript segment (so it never starts or ends mid-word), the other speaker is not talking across it, the handover is not overlapped, at least 1.0 s of speech, and at least 1.0 word per second of speech.

That is strict on purpose — 21,596 raw turns yield 4,774 clean. Spot-listening the ungated set turned up fragments cut mid-word and "pauses" that were really the other speaker holding the floor.

Caveats

transcript is approximate. SPRING-INX ships segment-level transcripts, so words are apportioned across VAD speech seconds rather than force-aligned. No word-level timings ship, and transcript must not be treated as one.

Turns ending in overlap are excluded. clean requires eot_gap >= 0, so turns the other speaker cut into are not here. They exist in the pipeline and can be rebuilt from it.

Trailing audio is capped at 5 s. For the 154 rows with a longer gap the final span is the capped length while eot_gap keeps the true value. All of those are eot_bench_eligible = False.

Licensing and citation

Derived from SPRING_INX Tamil R1, SPRING Lab, IIT Madras — CC BY 4.0 (arXiv:2310.14654). This derived dataset carries the same licence.

@article{tamileot2026,
  title  = {Tamil End-of-Turn Detection for Voice Agents},
  author = {Santhosh},
  journal = {arXiv preprint arXiv:2609.05631},
  year   = {2026}
}