Datasets:
license: other
license_name: hi-singers-research-license
license_link: LICENSE
pretty_name: Hi-Singers
task_categories:
- audio-to-video
language:
- en
- zh
tags:
- singing-head-synthesis
- audio-driven
- talking-head
- video-generation
size_categories:
- 100K<n<1M
extra_gated_prompt: >-
By requesting access you agree to use Hi-Singers for non-commercial academic
research only, not to redistribute the raw data, not to use it to identify,
surveil, or harm any individual, and to honor takedown requests. The data
originates from publicly available performances by online singing creators,
collected in accordance with source-platform Terms of Service. Full terms: see
the Access & Responsible Use Agreement and LICENSE.
extra_gated_fields:
Full name: text
Affiliation: text
Intended use: text
I agree to use this dataset for non-commercial academic research only: checkbox
I agree not to redistribute the raw data: checkbox
I agree to honor takedown / opt-out requests: checkbox
extra_gated_button_content: I agree to the terms
Hi-Singers
A Comprehensive High-Quality Dataset for Expressive Audio-Driven Singing Head Synthesis
⚠️ Ethics & Responsible Use. Hi-Singers is released for non-commercial academic research only. Access is gated and subject to a usage agreement. The data consists of publicly released performances by online singing creators, collected in accordance with the Terms of Service of the source platforms. We provide a takedown / opt-out channel — see the Ethics & Data Use section below.
Dataset Summary
Hi-Singers is the first large-scale, high-quality, in-the-wild 2D video dataset specifically tailored for singing head synthesis. It contains 29,608 video segments totaling approximately 170 hours, curated from raw footage through a rigorous automated + manual filtering pipeline that enforces thematic adherence (singing-centric), high-resolution rendering (1080p source), and stable, forward-facing motion. A dedicated evaluation benchmark balanced across languages (Chinese, English) and musical genres (Pop, Classical, Rock, Hip-hop, etc.) is included.
Supported Tasks
- Audio-driven singing head / portrait video generation
- Talking-head → singing-head domain adaptation
- Rhythmic synchronization evaluation (Beat Alignment Score, BAS)
Languages
English (en) and Chinese (zh).
Dataset Structure
| Field | Description |
|---|---|
video |
Cropped, forward-facing singing segment (512×512, 25 fps, H.264) |
audio |
Corresponding audio track (16 kHz) |
language |
en / zh |
genre |
Pop / Classical / Rock / Hip-hop / Other |
duration |
20–25 s per clip |
Statistics: roughly balanced English/Chinese; genre distribution dominated by Pop, reflecting the natural distribution of in-the-wild singing videos.
Data Collection & Processing
- Raw collection — ~1,200 GB / 4,500 h of front-facing singing videos of online singing creators from YouTube, TikTok (Douyin), and Bilibili (1080p+ only).
- Preprocessing — 25 fps, 16 kHz audio, H.264, sliced into segments.
- Automated filtering — face alignment & height (FaceAlignment), lip-sync (SyncNet), motion displacement & Euler-angle filtering, face quality (DSL-FIQA).
- Manual filtering — a panel of experienced engineers scored a representative subset on five SBF metrics (singing relevance, aesthetic quality, movement stability, articulation clarity, DSL-FIQA); a logistic-regression model mapped these to inclusion decisions across the full set.
Considerations & Limitations
- Languages: currently Chinese & English only. The pipeline is language-agnostic; additional languages are planned future work.
- Genre balance: Pop is dominant, mirroring real-world distribution. A genre-balanced test subset is provided to mitigate evaluation bias.
- Viewpoint: front-facing shots only (consistent with current singing-head synthesis settings); profile / multi-view is future work.
Ethics & Data Use
The dataset comprises publicly released, performance-oriented content by online singing creators, collected in accordance with the Terms of Service of the source platforms and intended solely for non-commercial academic research. This follows the established, peer-reviewed practice of widely-used web-sourced datasets such as CelebV-HQ (ECCV 2022), VoxCeleb (Interspeech), and VFHQ.
We take privacy and rights seriously:
- Gated access + a research-only usage agreement (see above).
- No sensitive-attribute annotation or inference.
- Takedown / opt-out: any individual appearing in the dataset, or rights holders, may request removal at any time via charleszhang@mail.ustc.edu.cn; we will promptly remove the relevant segments and re-issue the affected release.
- A Datasheet (see
DATASHEET.md) documents motivation, composition, collection, and recommended uses.