DiffSinger vs SoulX-Singer

DiffSinger is a free, open-source singing voice synthesis system built around a shallow diffusion mechanism. The maintained version adapts generated audio to 44.1 kHz and incorporates updated acoustic models plus faster diffusion sampling methods. Variance models and parameters provide control over pitch, energy, breathiness, and other vocal characteristics. The linked MakeDiffSinger project supplies dataset-building pipelines, including recording-slicing and labeling tools. Its guide describes preprocessing data, training models, and running inference on DS files with variance or acoustic models. Installation requires Python 3.10 or later, with PyTorch 2.4.0 or later recommended. ONNX is the deployment format. The repository links OpenUtau and DiffScope as production or deployment projects; DiffScope is an editor powered by DiffSinger and remains under development. Users prepare their own data and assets. The project prohibits generating another person’s voice without consent and includes a utility for removing speaker embeddings from checkpoints before distribution. Support channels include tutorials, GitHub issues and discussions, QQ, and Discord.

SoulX-Singer is a singing voice synthesis model that generates voices for singers it has not been tuned on. Pitch, rhythm, and expression can be guided by melody-conditioned F0 contours or score-conditioned MIDI notes. Its voice-conversion system changes raw singing audio to a target singer’s voice while retaining melody, rhythm, and lyrics, without requiring lyric or MIDI transcriptions. It supports Mandarin Chinese, English, and Cantonese, plus singer-timbre cloning, cross-lingual synthesis, and lyric editing while retaining natural prosody. Preprocessing tools handle vocal separation, dereverberation, F0 extraction, voice activity detection, and lyric and note transcription. Users can edit MIDI metadata, including lyrics, phoneme alignment, pitches, and durations, before synthesis. Soul-AILab provides generation and conversion demos and a separate MIDI editor through Hugging Face Spaces; the repository also supports local inference. Code and model weights use the Apache 2.0 license. Maintainers caution that automatic preprocessing may misalign singing with lyrics and notes, so manual correction may be needed.