Benson/seamless-segment-dataset

Seamless Interaction — Emotion Segments

This dataset is a derived, segmented version of the Seamless Interaction Dataset (Meta). Each row is one short clip (segment) of a single participant with aligned video, audio, and emotion labels.

---

Source dataset

• Name: Seamless Interaction Dataset
• Provider: Meta Platforms, Inc. (FAIR Seamless Team)
• Content: Large-scale face-to-face interactions (video, audio, transcripts, VAD, movement/emotion features from the Imitator model)
• Original format: Per-participant files (e.g. one .mp4, .wav, .npz, .json per file_id) organized by label/split/batch/archive
• HuggingFace (original): facebook/seamless-interaction

Only participants with imitator movement features (has_imitator_movement=1 in the filelist) are used.

---

How this dataset was created

1. Input: The official filelist (e.g. assets/filelist.csv) is filtered to rows with has_imitator_movement == 1. Each such row corresponds to one participant (file_id).

2. Per participant (script: download_and_segment_participant.py):

   • The participant’s data is downloaded from the source (S3) into a local directory.
   • Emotion runs: Per-frame dominant emotion is computed as argmax(movement:emotion_scores) (8 classes: Anger, Contempt, Disgust, Fear, Happiness, Neutral, Sadness, Surprise). Contiguous frames with the same dominant are grouped into runs.
   • Segmentation: Runs are split into segments of 3–10 seconds (by FPS). Only segments where the participant is speaking are kept:
     • At least 50% of the segment duration must overlap with speech intervals (from metadata:vad and metadata:transcript).
     • At least 3 seconds of continuous speech (with up to 2 s gaps merged as “thinking pause”) must fall inside the segment.
   • Segment label: For each segment, the segment-level dominant emotion is not the run label; it is argmax(mean(emotion_scores[start_frame:end_frame], axis=0)) — i.e. average emotion scores over the segment, then argmax.
   • Extraction: For each segment, video and audio are cut with ffmpeg (output seek + re-encode for video so the clip is exactly the requested time window; audio copied). Per-segment metadata and emotion-related arrays (valence, arousal, emotion_scores, etc.) are written to the segment folder. Original full-length .mp4, .wav, and .npz are deleted after segmenting; only segment subfolders are kept.

3. Upload (script: batch_segment_and_upload_hf.py):

   • The batch script iterates over all filelist participants (with the above filter), runs the segment script for each, collects all segment rows (video, audio, metadata, segment_data), and pushes to this HuggingFace dataset repo.
   • Progress is logged locally so runs can be resumed (already-processed participants are skipped).
   • The dataset is pushed every N successful participants (e.g. every 2). New segment rows are concatenated with any existing data on the Hub before each push.
   • Dataset card: This README is the dataset card; it is uploaded to the repo when the batch script pushes.

Parameters used in the segmentation (defaults):

• Segment length: 3–10 s (by FPS).
• Min speech fraction: 0.5 (50% of segment must be speech).
• Min continuous speech: 3 s (within segment, with 2 s merge gap for pauses).
• VAD merge gap: 2 s (gaps ≤ 2 s merged into one “turn”; larger gaps split).

---

Dataset structure

Each row in this dataset is one segment (one clip of one participant).

Column	Type	Description
video	Video	Segment video (MP4), same time window as audio.
audio	Audio	Segment audio (WAV), 48 kHz.
file_id	string	Original participant ID (e.g. V00_S0809_I00000309_P0947).
segment_id	string	Segment name (e.g. segment_000).
label	string	Dataset label: improvised or naturalistic.
split	string	Split: train, dev, or test.
batch_idx	int	Batch index from the original layout.
archive_idx	int	Archive index from the original layout.
metadata	string	JSON string: start_frame, end_frame, start_time, end_time, duration_seconds, dominant_emotion_index, dominant_emotion_name.
segment_data	string	JSON string: full emotion-related data for the segment (dominant index/name, emotion_scores, valence, arousal, tokens, etc.; see original movement features).

Emotion classes (index 0–7): Anger, Contempt, Disgust, Fear, Happiness, Neutral, Sadness, Surprise.

---

Intended use

• Training or evaluating models that use short, emotion-labeled audiovisual clips (e.g. emotion recognition, multimodal fusion).
• Studying affect over 3–10 s speaking segments with aligned video, audio, and continuous emotion/valence/arousal.

---

License and attribution

This derived dataset follows the license and terms of the original Seamless Interaction Dataset (e.g. CC-BY-NC 4.0 or as specified by Meta). You must comply with the original dataset’s license and usage policy.

Attribution: Segments are derived from the Seamless Interaction Dataset by Meta Platforms, Inc. Processing and segmentation were done with the open-source seamless_interaction scripts (download_and_segment_participant.py, batch_segment_and_upload_hf.py).

---

How to load

from datasets import load_dataset

ds = load_dataset("YOUR_USERNAME/seamless-segment-dataset", split="train")
# Each row: video, audio (48 kHz), file_id, segment_id, label, split, batch_idx, archive_idx, metadata (JSON string), segment_data (JSON string)

Parse metadata and segment_data with json.loads(row["metadata"]) and json.loads(row["segment_data"]) for numeric and emotion fields.