Shofo Talking Head En: 10,000 Short-Form Social Videos with Transcripts

Shofo Talking Head Dataset (English) contains approximately 10,000 talking-head videos totaling about 186 hours, with a mean clip length of 67 seconds. It was built and released by Shofo using a multi-stage curation pipeline for clean single-speaker framing. The dataset was last updated on July 1, 2026.

Use Cases

• Train avatar generation models based on curated single-speaker video clips.
• Develop lip-sync algorithms based on time-aligned transcripts paired with video.
• Improve text-to-speech and dubbing systems based on the dataset's clean speech and framing.
• Benchmark computer vision models for face and framing analysis on filtered social video content.

Strengths

• Approximately 10,000 video clips provide a substantial volume of data.
• Videos are filtered for clean single-speaker framing, suggesting high visual quality for the task.
• Each clip is paired with time-aligned transcripts, enabling precise audio-visual alignment.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

Shofo

Collection Method

Built via a multi-stage pipeline involving face/framing analysis, on-screen-text detection, object-occlusion detection, and voice/face matching.

Freshness

Last updated 2026-07-01 02:44:43; freshness should be verified.