MeghanaKap/podcast-interactivity

Podcast Interactivity Clips

Short clips of English conversational speech between two people, each centred on one interaction moment (a pause, a turn change, a backchannel or an interruption), for reinforcement-learning a full-duplex speech model's timing. Built to extend the Seamless Interaction recipe of Ohashi et al. 2026, Multi-Faceted Interactivity Alignment in Full-Duplex Speech Models, which produced kyutai/personaplex-rl-seamless.

As of 2026-10-30: 16,263 clips from 219 episodes in 189 shards (160.4 h of clip audio including context).

Research use only. Access is gated and granted by hand; do not redistribute the audio.

What a clip is

Each clip is the user side of a moment in a conversation: the audio a full-duplex model would hear from the person it is talking to. The other speaker (the reference listener) is not in the audio; their timing is kept as labels, so a model playing that listener can be scored against what the real person did.

Axis	Clips	Share	Clip rule
Pause	11,652	72%	One user utterance of at least 4 s with an internal pause; the other speaker silent
Turn-taking	1,798	11%	User finishes; the other speaker replies within 0.4 s (clip at least 5 s)
Backchannel	1,982	12%	User utterance of at least 5 s; the other speaker only says short things (at most 1 s), at least once
Interruption	831	5%	User, other, user, other, with the user cutting in before the other finishes (clip at least 3 s)

Clips carry up to 30 s of the conversation before the moment (median 30 s), so training can sample context. The moment itself has a median length of 10.9 s (5th to 95th percentile 4.8 to 26.1 s).

Sources

Channel	Episodes	Clips
Raj Shamani	190	12,976
Ranveer Allahbadia	29	3,287

Episodes come from a per-channel list of every upload (videos.csv), filtered to English originals of 10 to 240 minutes, newest first. Episodes with dubbed audio tracks in other languages are processed from their original-language track only.

How it was built

1. Language: the original audio's language from the video metadata, confirmed with Whisper language ID on three 30 s windows. English only.
2. Speakers: pyannote speaker-diarization-3.1. An episode is kept only if its two main speakers hold at least 90% of the speech; clips touching any third voice are dropped.
3. Speech boundaries: each speaker's diarized turns are trimmed to Silero VAD speech, so pauses and gaps are measured on real silence. Utterances are speech joined across silences of at most 1.0 s.
4. Clips: the four rules above, applied with each speaker in turn as the user. The first and last 5 minutes of every episode (intros, ads, outros) are skipped.
5. User track: 16,263 clips are the episode mix with only the user's diarized speech kept and everything else silenced; 0 come from episodes recorded with each speaker on their own stereo channel, where the user's own channel is used as is.
6. Transcripts: Whisper large-v3-turbo on each of the user's utterances, for an LLM judge's conversation history.

Format

WebDataset tar shards in data/shard-NNNNN.tar, each holding one or more whole episodes, each with an index in index/shard-NNNNN.jsonl (one line per clip, the same metadata as in the tar).

File in the tar	Content
<key>.flac	User track, 24 kHz mono, 16-bit, from the start of the available context to the end of the moment
<key>.json	Metadata, below

Field	Meaning
key	<video id>_<axis>_<speaker>_<ms>
axis	pause, turn, backchannel or interrupt
start, end	The moment, in seconds from the start of the audio (start = context available)
t_ref	Turn-taking: when the user's turn ends; interruption: when the user's interruption ends (seconds from start)
ref_bc	Backchannel: start times of the real listener's backchannels (seconds from start)
transcript	The user's utterances in the clip: start, end (seconds from the audio start), text
leak_s	Seconds where the other speaker overlaps the user in the user track (0 for stereo-channel clips)
track	diarized mix or stereo channel
video, channel, title, speaker	Source episode, and which of its two main speakers is the user
episode_gaps	The episode's turn-gap statistics (and the stereo split levels when used)

Loading

With the training code in s2s/personaplex_rl:

personaplex_rl/.venv/bin/python personaplex_rl/train/train.py --data hf://MeghanaKap/podcast-interactivity

Or directly:

import io, json, tarfile
import soundfile as sf
from huggingface_hub import hf_hub_download

repo = "MeghanaKap/podcast-interactivity"
index = [json.loads(l) for l in open(hf_hub_download(repo, "index/shard-00000.jsonl", repo_type="dataset"))]
clip = index[0]
with tarfile.open(hf_hub_download(repo, clip["shard"], repo_type="dataset")) as tar:
    audio, sr = sf.read(io.BytesIO(tar.extractfile(clip["key"] + ".flac").read()))
moment = audio[int(clip["start"] * sr):int(clip["end"] * sr)]

Known limitations

• Overlap leak: in diarized mix clips the other speaker leaks into the user track wherever both talk at once. Pause and turn-taking clips with more than 0.5 s of leak were dropped; backchannel and interruption clips keep their overlap, because overlap is what defines them. Median leak 0.00 s, 95th percentile 1.18 s.
• Edited episodes: many podcasts cut pauses in editing, which makes turn changes unnaturally tight. Use episode_gaps (median_gap_s, tight_gap_frac, overlap_frac) to filter or weight episodes.
• Diarization errors: speaker turns come from pyannote, not separate microphones, so short backchannels in overlap can be missed or given to the wrong speaker.
• Narrow sources: a few Indian English-language interview podcasts; the speech style and the hosts repeat across many clips.
• Transcripts are automatic (Whisper) and unreviewed.