mavleo96/turn-detection-data

Turn Detection Data

End-of-turn (EoT) training data for a voice agent. Every time a voice activity detector (VAD) says a speaker stopped talking, we take that speaker's audio from the start of their turn up to that moment and label it:

• complete: the other person takes the floor next
• incomplete: the speaker pauses and goes on

No audio after the stop is included (it would give the label away). Clips are 16 kHz mono FLAC, at most 30 s.

Sources

source	type	train	validation	test
SpokenWOZ	phone calls (task-oriented)	167,376	20,510	40,353
Seamless Interaction (naturalistic)	in-person conversations	165,908	17,090	20,221
total		333,284	37,600	60,574

About half the rows are complete (50.4% train, 49.8% validation, 52.0% test). Splits follow the sources' own splits: no dialogue or (for Seamless) person appears in more than one split. Labels come from SpokenWOZ's annotated turns and from Seamless's word timestamps.

Columns

column
id	{source}_{conversation}_{channel}_{k:03d}
audio	16 kHz mono FLAC
data_source	spokenwoz / seamless
original_data_id	dialogue / conversation id in the source
channel	SpokenWOZ: 0 = user, 1 = system. Seamless: the two participants
timestamp_start, timestamp_end	clip span in the conversation (s); end = when the VAD stop fires
label	incomplete / complete
meta	JSON string with source-specific details (transcript text, gaps, turn info)

from datasets import load_dataset
ds = load_dataset("mavleo96/turn-detection-data")
phone = ds["train"].filter(lambda x: x == "spokenwoz", input_columns="data_source")

dropped/ (points left out, with the reason), stats/ and report/ hold build details per source and split.

How it was made

See GENERATE.md for the full method, settings and commands; code is in scripts/.

License

CC BY-NC 4.0 (non-commercial), following both sources: SpokenWOZ and Seamless Interaction are CC BY-NC 4.0. Please cite the original datasets when using this data.