SingNet: Towards a Large-Scale, Diverse, and In-the-Wild Singing Voice Dataset

TL;DR

SingNet tackles the scarcity of large-scale, diverse singing data by introducing an automated data-processing pipeline that harvests in-the-wild singing data, yielding about 3000 hours across multiple languages and styles. The authors pre-train and open-source state-of-the-art models (Wav2vec2, BigVGAN, NSF-HiFiGAN) on SingNet and benchmark Automatic Lyric Transcription, neural vocoding, and Singing Voice Conversion, providing actionable baselines and resources for SVS/SVC research. Results show that large-scale singing data enhances ALT performance directly, improves vocoder quality for singing contexts, and enables stronger zero-shot SVC as dataset size grows. The work offers a scalable, extensible framework and public data to improve reproducibility and practical deployment in singing voice technologies.

Abstract

The lack of a publicly-available large-scale and diverse dataset has long been a significant bottleneck for singing voice applications like Singing Voice Synthesis (SVS) and Singing Voice Conversion (SVC). To tackle this problem, we present SingNet, an extensive, diverse, and in-the-wild singing voice dataset. Specifically, we propose a data processing pipeline to extract ready-to-use training data from sample packs and songs on the internet, forming 3000 hours of singing voices in various languages and styles. Furthermore, to facilitate the use and demonstrate the effectiveness of SingNet, we pre-train and open-source various state-of-the-art (SOTA) models on Wav2vec2, BigVGAN, and NSF-HiFiGAN based on our collected singing voice data. We also conduct benchmark experiments on Automatic Lyric Transcription (ALT), Neural Vocoder, and Singing Voice Conversion (SVC). Audio demos are available at: https://singnet-dataset.github.io/.

Figures (7)

• Figure 1: An overview of the SingNet data processing pipeline. It processes in-the-wild songs and sample packs into a ready-to-use dataset for model training.
• Figure 2: Duration statistics (hours) of SingNet by language and style sorted by the data scales. "MIS" means uncoded Indigenous languages
• Figure 3: BPM and Pitch statistics (occurrences) of SingNet. The pitch is illustrated as MIDI notes where A4=69=440Hz. Values outside of the illustrated ranges are considered errors and are removed .
• Figure 4: Comparison of acoustic and semantic diversities between SingNet and the mixture of existing datasets. It can be observed that SingNet has more diverse data regarding both semantic and acoustic levels than the mixture of existing datasets.
• Figure 5: An overview of the audio annotation website. The sample packs used in annotation, the annotator, and the author of the annotation system are all made to be anonymous.