ProgramComputer/avspeech-visual-audio

AVSpeech Video + Audio

This repository is a media-bearing reconstruction of the public AVSpeech annotations. Each row represents an already-trimmed segment and keeps the original source-video timing and target-face-center metadata.

Dataset structure

• clip_id: identifier derived as {youtube_id}_{start_sec:.3f}_{end_sec:.3f}.
• avspeech_metadata: JSON containing youtube_id, start_sec, end_sec, x_center, and y_center from the AVSpeech annotation.
• video: video-only stream, or null when the source segment could not be materialized.
• audio: audio-only stream, or null when the source segment could not be materialized.

The video field is already trimmed. start_sec and end_sec refer to the original YouTube-video timeline and must not be used to seek again within this clip. AVSpeech defines (x_center, y_center) as the normalized center of the speaker's face in the frame at the beginning of the segment, with (0, 0) at the top left.

Audited snapshot and known publication gap

The following figures describe revision efdceb2a0b9d81a6aec76f10668cca49e8209e37:

Split	Rows	Rows with both media	Rows without both media	Parquet files	Encoded size
train	2,621,845	1,589,842	1,032,003	5,142	1,404,473,032,806 bytes
test	183,273	98,605	84,668	359	88,861,772,006 bytes
total	2,805,118	1,688,447	1,116,671	5,501	1,493,334,804,812 bytes

The train states were verified by an exhaustive read-only scan of all 5,142 train Parquet files and 2,621,845 rows. Of the 1,032,003 train rows without both media streams, 1,032,000 have both media null, 3 are video-only, and 0 are audio-only. The test split was not audited at that row-state granularity, so the table reports only its aggregate count without both streams.

The completed exporter expected 1,589,942 paired train occurrences, while this snapshot contains 1,589,842, an aggregate gap of 100. Surviving non-media evidence supports high-confidence assignment of 49 of those occurrence slots:

• 46 both-null occurrences have a paired sibling at the pinned revision.
• 3 video-only occurrences retain a published video.path while audio.path is null, showing that the exporter reached archive-member processing.

The remaining 51 occurrence slots cannot be assigned to exact rows or clip_id values without the original expected-pair manifest or historical ID-to-archive map. Their ambiguity remains within a pool of 4,103 metadata-resolved both-null occurrences across 403 YouTube IDs. “Unresolved” does not mean that these rows were verified unavailable.

The audit selected only clip_id, avspeech_metadata, video.path, and audio.path. It did not select or materialize embedded media bytes, and its range guards recorded zero intersections with media-byte column chunks. No media recovery, YouTube retrieval, torrent-media transfer, row repair, or Hugging Face mutation was attempted as part of that audit.

Bounded loading

Do not use snapshot_download for routine training: the repository is about 1.49 TB. Stream rows, keep media decoding disabled at the dataset layer, skip any row without both media streams, and materialize only one bounded work unit at a time. Preserve the official split and stable row provenance, and report pre-filter and retained denominators plus exclusions by reason.

from datasets import Audio, Video, load_dataset

revision = "efdceb2a0b9d81a6aec76f10668cca49e8209e37"
rows = load_dataset(
    "ProgramComputer/avspeech-visual-audio",
    split="train",
    revision=revision,
    streaming=True,
)
rows = rows.cast_column("video", Video(decode=False))
rows = rows.cast_column("audio", Audio(decode=False))

for row in rows:
    if row["video"] is None or row["audio"] is None:
        continue
    # Materialize/process this row in bounded temporary storage.

The official AVSpeech page states that its supplied train and test annotations use disjoint speakers. This reconstruction preserves those source split labels. It does not add person identities, and a YouTube video ID must not be described as a speaker identity.

Intended use and limitations

This dataset is intended for research on audio-visual speech and related representation-learning tasks. It is derived from public Internet video and is not demographically balanced. Availability, codecs, media quality, language, pose, lighting, and annotation accuracy vary. Missing rows are not necessarily random, so filtering to paired media may introduce additional selection bias. The unresolved 51-slot publication gap is aggregate provenance information, not a verified unavailable-row list, and must not be converted into invented row-level labels.

The face-center coordinate is a point hint at the beginning of the segment, not a bounding box, persistent track, verified identity label, or consent signal. Downstream systems must validate the associated detected face and must not use this dataset for identification, surveillance, or consequential decisions.

License and provenance review

The official AVSpeech download page provides train/test annotation CSVs and states that “this data” is available under CC BY 4.0. This repository also redistributes media derived from YouTube videos. The maintainer has not yet documented a legal review establishing that the same license statement covers redistribution of every embedded media stream or that all upstream platform and uploader terms are satisfied. Therefore this card deliberately does not assert a Hugging Face license tag for the media-bearing reconstruction.

Before continued public redistribution, document the source acquisition process, takedown procedure, upstream terms, and the basis for redistributing the embedded audio/video. This note is a publication safeguard, not legal advice.

Citation

If you use the data, cite the original AVSpeech work:

@article{ephrat2018looking,
  title={Looking to Listen at the Cocktail Party: A Speaker-Independent Audio-Visual Model for Speech Separation},
  author={Ephrat, Ariel and Mosseri, Inbar and Lang, Oran and Dekel, Tali and Wilson, Kevin and Hassidim, Avinatan and Freeman, William T. and Rubinstein, Michael},
  journal={ACM Transactions on Graphics},
  year={2018}
}

Official AVSpeech project and download page: https://looking-to-listen.github.io/avspeech/download.html.