---
license: other
task_categories:
  - video-classification
  - text-to-speech
  - automatic-speech-recognition
language:
  - en
tags:
  - talking-head
  - video
  - lip-sync
  - avatar
  - multimodal
  - social-media
  - vertical-video
size_categories:
  - 1K<n<10K
pretty_name: Shofo Talking Head (English)
extra_gated_heading: "Request access to the Shofo Talking Head Dataset"
extra_gated_prompt: "A curated ~10,000-clip sample of social talking-head video with timestamped transcripts. Shofo has aggregated billions of videos into a searchable index to curate custom datasets at scale. Fill out the following for access:"
extra_gated_fields:
  Full name: text
  Work email: text
  Company: text
extra_gated_button_content: "Request access"
---

# Shofo Talking Head Dataset (English)

A curated dataset of **~10,000 talking-head videos** (~186 hours, mean ~67s/clip), filtered for clean
single-speaker framing and paired with time-aligned transcripts. Built and released by **[Shofo](https://shofo.ai)**.

This dataset is short-form social video, designed to support modern avatar, lip-sync, dubbing, and TTS work.
Every clip is curated by a multi-stage pipeline (face/framing analysis, on-screen-text detection,
object-occlusion detection, voice/face matching, language detection) to provide a curated dataset rather than a raw dump.

<p align="center"><img src="https://huggingface.co/datasets/Shofo/talking-head-assets/resolve/main/talking_head_grid.png" width="520" alt="Example clips from the dataset"></p>

## What makes a clip eligible
Each clip in this release satisfies all of the following (derived by Shofo's curation pipeline):
- **Single, centered, persistently-visible face** (face present ≥95% of frames, low framing jitter, no multi-face).
- **On-camera speech**: the speaker's voice matches the on-screen face (not dubbed/voiceover).
- **Audio-visual sync verified**: lip motion correlates with the audio, checked with a SyncNet-based AV-sync model plus a sound-attribution pass that flags clips mimed to a borrowed sound. Lip-syncs, dubs, and mimed clips are removed.
- **Active speaking** (detected mouth motion).
- **Face not occluded** (screened for objects blocking the face; occlusion of the body or by props is not fully filtered).
- **English** (dual-detector language ID, high confidence).
- **Duration 10s to 300s.**

The ~10,000 clips are selected by a composite quality score together with the curation gates above.

## Dataset structure
```
metadata.parquet            # one row per clip (schema below)
videos/{ab}/{video_id}.mp4   # H.264/MP4, sharded by first 2 chars of id   [if files hosted]
```

### Fields
| field | description |
|---|---|
| `video` | the clip itself (decoded Video feature, auto-generated) |
| `video_id` | TikTok numeric video id (stable reference) |
| `tiktok_url` | handle-less source link `https://www.tiktok.com/@i/video/{id}` |
| `duration_ms` | clip duration (ms) |
| `resolution`, `width`, `height` | frame dimensions |
| `fps`, `codec`, `bitrate`, `has_audio` | technical metadata (ffprobe) |
| `language` | detected spoken language (en) |
| `has_music` | music detected in audio |
| `transcript` | time-aligned, segment-level **WEBVTT** transcript with punctuation (all clips) |

Every clip is selected by Shofo's multi-signal curation pipeline (face/framing, on-screen-text,
occlusion, voice/face matching, language); see *What makes a clip eligible* above.

Transcripts cover all clips and were produced with **NVIDIA Canary-Qwen-2.5B**.

## Usage
```python
from datasets import load_dataset

# Auto-detected as a videofolder: each row has a decoded `video` column
# plus all metadata fields (scores, transcript, technical info).
ds = load_dataset("Shofo/shofo-talking-head-en", split="train")
row = ds[0]
print(row["video_id"], row["resolution"], row["duration_ms"])
video = row["video"]   # decoded video (Video feature)
```
Each row also carries `file_name` if you prefer to fetch the raw mp4 directly via `huggingface_hub.hf_hub_download`.

## Known limitations
- **Resolution is modest.** Clips are vertical social video, 88% 576×1024 (SD) and 9% 720×1280; curation and
  composition quality is high but native resolution is not HD. Face crops are typically 300 to 400px tall.
  (All clips are H.264/MP4, portrait, with audio, bitrate 400 kbps or higher.)
- **A small number of lip-syncs may remain.** Clips were screened for audio-visual sync (SyncNet-based scoring
  plus a sound-attribution check) and flagged lip-syncs, dubs, and mimed clips were removed, but a low
  single-digit percentage of hard cases can slip through. Screen for it if your task is sync-sensitive.
- **Some fine-grained issues remain.** Composition signals can't catch everything (e.g. green-screen compositing).

## About Shofo & more data
This is a curated sample. Shofo has aggregated **billions** of social videos into a searchable index with rich
derived metadata (captions, embeddings, quality/occlusion/text scores, transcripts), and builds custom curated
datasets at scale. **Interested in larger or task-specific data? Reach us at founders@shofo.ai / https://shofo.ai**