xiaoyu1104/talkdataset

TalkVid processed video clips

This repository is a processed mirror/subset in progress of FreedomIntelligence/TalkVid, not a new original dataset. Credit belongs to the TalkVid authors. See the official project and paper. Dataset use is subject to CC BY-NC 4.0 and the original project's non-commercial research terms.

Files and progress

Files are uploaded incrementally. The original metadata is divided into five logical batches of approximately 20% of records each. A batch directory existing does not mean it is complete. A source can be unavailable; errors must not be interpreted as successful samples. Only uploaded shards are available here.

• data/batch-01/ through data/batch-05/: TAR shards, normally up to 1,000 samples or about 4 GiB of media per shard (a single large sample may exceed this target).
• metadata/batch-XX/*.jsonl: the corresponding sample annotations.
• Each TAR contains <Person ID>_<original id>.mp4, <Person ID>_<original id>.m4a, and <Person ID>_<original id>.json. The index is zero-based in the pinned metadata order.

Processing and synchronization

Videos are downloaded from the YouTube URLs in the original annotations. Source resolution is retained. Video is encoded as H.264 CRF 18; audio as AAC 128 kbps. Both streams are cut on one shared source time axis. Start and end cuts are snapped forward to actual source video frame boundaries, and the same origin is subtracted from audio and video. actual_source_start and actual_source_end record the effective cuts; original time/frame annotations remain in the JSON. No independent video time shift is applied.

Checks require a non-black first decoded frame, near-zero stream start times, source-matching video frame times, and decoded audio alignment with the same source interval (within 1 ms for non-silent clips). Independent M4A audio matches the MP4 audio. These checks preserve the source's timing; they do not establish that every original source has perfect perceptual lip synchronization. Problematic clips are recorded locally for review instead of silently treated as successful.

Each shard is verified against its immutable Hub commit and SHA-256 before its local copy is removed. Full-resolution media can be large; this repository is a work in progress and may pause due to source availability or storage quotas.

use_id in each JSON is copied from the original info["Person ID"]. Example: 597_videovideoTr6MMsoWAog-scene1-scene1.mp4. A source ending may produce a smaller shard; the limit is a maximum, not a fixed shard size.

Scheduling prioritizes Person ID groups, but idle workers may start later people while slow jobs for an earlier person continue. At most six source jobs run concurrently, with at most two ffmpeg/ffprobe processes at a time to bound CPU and memory use. Per-source failures remain explicitly incomplete; upload verification, disk or proxy failures pause the run. Logical batch paths retain their original index-based partition and may populate out of order.

Source frame timestamps are inspected only around each target interval using bounded local scans, rather than decoding each full source in advance. Original source timestamps and the shared audio/video origin remain unchanged.