You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

By requesting access, you agree to use Hi-Singers for non-commercial academic research only; not to redistribute the raw data; not to identify, surveil, or harm any individual; and to honor takedown requests. The data originates from publicly released performances by online singing creators and was collected in accordance with source-platform Terms of Service. Full terms are available in the Hi-Singers Research License (LICENSE.txt).

Log in or Sign Up to review the conditions and access this dataset content.

Hi-Singers

A Comprehensive High-Quality Dataset for Expressive Audio-Driven Singing Head Synthesis

Paper Clips Duration Languages License Access

Paper / DOI · Supplementary material · Browse files · License · Citation

Research use only. Hi-Singers is a gated dataset released for non-commercial academic research. Access requires acceptance of the Hi-Singers Research License. Redistribution of the raw data and attempts to identify, surveil, or harm individuals are prohibited.

The paper and dataset have different terms: the paper is published under CC BY-NC-ND 4.0, while the dataset is governed by the separate Hi-Singers Research License.

Hi-Singers three-stage data construction pipeline
Figure 1. The Hi-Singers construction pipeline: data collection and preprocessing, automated filtering, and human-centered filtering.

Overview

Hi-Singers is a large-scale, high-quality, in-the-wild 2D video dataset designed specifically for expressive audio-driven singing-head synthesis. It addresses the domain gap between speech and singing: singing involves stronger articulation, larger jaw motion, richer facial expression, musical accompaniment, and strict rhythmic synchronization.

The release contains 29,608 curated video segments totaling approximately 170 hours. Samples are selected through an automated and human-reviewed pipeline that emphasizes singing relevance, facial quality, lip-sync consistency, stable forward-facing motion, and expressive clarity. A dedicated 200-clip evaluation benchmark, balanced across language and musical style, supports reproducible comparison.

At a glance

Hi-Singers
Scale 29,608 clips · approximately 170 hours
Clip duration 20-25 seconds
Video 512 × 512 · 25 fps · H.264
Audio 16 kHz
Languages English and Chinese, approximately balanced
Genres Pop, Classical, Rock, Hip-hop, and Other
Capture setting In the wild · front-facing singing performances
Source quality Source videos selected at 1080p or higher
Evaluation set 200 clips balanced across language and musical style

Representative Hi-Singers samples from Pop, Classical, Hip-hop, and Rock
Figure 2. Representative expressive singing clips spanning Pop, Classical, Hip-hop, and Rock.

Hi-Singers language and music genre distributions
Figure 3. Language and musical-genre distributions across the 29,608 released clips.

Why Hi-Singers?

  • Singing-specific: centered on genuine singing rather than speech or background-music-only segments.
  • Expressive and photorealistic: preserves vivid articulation, large jaw openings, and high-quality 2D facial texture.
  • Rhythm-aware: supports evaluation with the Beat Alignment Score (BAS), which measures alignment between audio beats and facial-motion peaks.
  • Carefully curated: combines face resolution and alignment checks, SyncNet lip-sync filtering, motion and pose filtering, DSL-FIQA quality assessment, and human scoring.
  • Benchmark-ready: includes controlled evaluation across English and Chinese as well as multiple musical genres.

Dataset construction

The paper's processing pipeline has three stages:

  1. Collection and preprocessing - approximately 4,500 hours / 1,200 GB of candidate footage; 1080p+ source selection; conversion to 25 fps, 16 kHz audio, and H.264; temporal segmentation.
  2. Automated filtering - face size and alignment, active-speaker / lip-sync checks, displacement and Euler-angle motion constraints, and facial-quality scoring.
  3. Human-centered filtering - multi-annotator assessment of singing relevance, visual quality, motion stability, articulation clarity, and facial quality, followed by score fusion.

Data retention through the Hi-Singers filtering stages
Figure 4. Data retention through successive filtering stages, from raw candidates to the final curated release.

The resulting release retains approximately 6.9% of the raw candidate data.

Data organization

The dataset is distributed as gated ZIP shards under Hi-singers/. Each sample is organized around a cropped, forward-facing singing clip and its corresponding audio, language, genre, and duration information.

Field Description
video Cropped forward-facing singing segment, 512 × 512, 25 fps, H.264
audio Corresponding 16 kHz audio
language en or zh
genre Pop, Classical, Rock, Hip-hop, or Other
duration 20-25 seconds per clip

Pop is the largest genre, reflecting the natural distribution of in-the-wild singing videos. The evaluation subset is balanced to reduce genre-related comparison bias.

Access and download

  1. Sign in to Hugging Face and accept the access terms at the top of this page.
  2. Authenticate locally with a Hugging Face account that has been granted access:
hf auth login
  1. Download the complete dataset:
hf download CharlesZhang-USTC/Hi-Singers \
  --repo-type dataset \
  --local-dir ./Hi-Singers

Or download only the archive shards:

hf download CharlesZhang-USTC/Hi-Singers \
  --repo-type dataset \
  --include "Hi-singers/*.zip" \
  --local-dir ./Hi-Singers

Python users can use huggingface_hub:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="CharlesZhang-USTC/Hi-Singers",
    repo_type="dataset",
    local_dir="./Hi-Singers",
)

Benchmark and paper highlights

The paper evaluates 3D-coefficient and diffusion-based singing-head models. Training Hallo on Hi-Singers improves all reported dimensions over the tested generic-data baselines, including visual fidelity, lip synchronization, rhythmic alignment, and mean opinion score. The reported Hi-Singers-trained result reaches FID 23.15, FVD 181.42, Sync-C 7.72, BAS 0.168, and MOS 3.92 on the singing-head benchmark.

For the full protocol, ablations, cross-language breakdown, and cross-genre results, see the paper / DOI record and supplementary material.

Intended uses

Hi-Singers is intended for non-commercial academic research in areas such as:

  • audio-driven singing-head or portrait-video generation;
  • talking-head to singing-head domain adaptation;
  • expressive facial-motion modeling;
  • lip-sync and rhythmic-synchronization evaluation;
  • cross-language and cross-genre analysis of singing-head synthesis.

The dataset must not be used for identity inference, surveillance, impersonation, harassment, harmful synthetic media, or any purpose prohibited by the license.

Limitations

  • Languages: the current release covers English and Chinese.
  • Genre distribution: Pop is naturally over-represented in the full dataset.
  • Viewpoint: clips emphasize stable, forward-facing faces; profile and multi-view coverage is limited.
  • Provenance: the data is sourced from publicly released online performances and may inherit platform, cultural, demographic, and creator-popularity biases.
  • Scope: strong benchmark results do not guarantee safe or reliable deployment outside the research settings described in the paper.

Ethics, privacy, and takedown

Hi-Singers contains publicly released, performance-oriented content by online singing creators. The dataset is gated, limited to non-commercial academic research, and does not provide sensitive-attribute annotations. Researchers are responsible for complying with the license, applicable law, institutional review requirements, and source-platform terms.

Individuals appearing in the dataset, or relevant rights holders, may request removal at any time. Contact harryjun@ustc.edu.cn with enough information to identify the affected material. Valid requests will be reviewed and the relevant segments removed from a subsequent release.

Citation

If Hi-Singers is useful in your research, please cite the ACM Multimedia 2026 paper:

@inproceedings{zhang2026hisingers,
  author    = {Yichi Zhang and Hui Zhang and Guanjun Liu and Yuefeng Zou and Fengzhao Sun and Jun Yu},
  title     = {Hi-Singers: A Comprehensive High-Quality Dataset for Expressive Audio-Driven Singing Head Synthesis},
  booktitle = {Proceedings of the 34th ACM International Conference on Multimedia},
  series    = {MM '26},
  year      = {2026},
  month     = nov,
  location  = {Rio de Janeiro, Brazil},
  publisher = {Association for Computing Machinery},
  address   = {New York, NY, USA},
  doi       = {10.1145/3767308.3838646},
  url       = {https://doi.org/10.1145/3767308.3838646}
}

Authors and contact

Yichi Zhang*, Hui Zhang*, Guanjun Liu, Yuefeng Zou, Fengzhao Sun, and Jun Yu†
University of Science and Technology of China
* Equal contribution · † Corresponding author

For dataset questions or access issues, contact charleszhang@mail.ustc.edu.cn. For takedown requests, contact harryjun@ustc.edu.cn as specified in the dataset license.


Built for research on expressive, rhythm-aware, and high-fidelity singing-head synthesis.

Downloads last month
191