gonnerthetooner/orislop-ami-lipsync-data

Orislop AMI lip-sync source data

This repository mirrors an acquisition-capped subset of the official AMI Meeting Corpus for reproducible Orislop audio-visual synchronization research.

It contains:

• low-size AMI close-up AVI video streams;
• corresponding individual headset WAV streams;
• SHA-256 provenance records;
• the official camera/headset mapping snapshot and generated clip manifests after local acquisition completes.

Files are uploaded only after a local download finishes and is atomically renamed. In-progress .partial files are excluded.

License and attribution

AMI publishes its signals and transcription under CC BY 4.0. See AMI_ATTRIBUTION.md and the per-file provenance.jsonl records.

Official documentation:

• download page
• audio/video signal mappings
• ethics and consent

Important labeling boundary

AMI provides authentic synchronized audiovisual recordings. It does not provide truthful deepfake labels. Controlled timing offsets generated by the Orislop manifest pipeline are synchronization hard negatives and are labeled joint_forgery=0.

Predictions trained from this material are probabilistic synchronization signals, not proof that a person lied or created deceptive media.

Colab use

Use huggingface_hub.snapshot_download with allow_patterns to fetch only the meetings needed for a run. Avoid copying the entire dataset into Colab's ephemeral VM unless the selected runtime has sufficient disk.