SoulX-Singer Models (Safetensors Mirror)

Safetensors conversion of Soul-AILab/SoulX-Singer weights for use in the MAESTRO AI Workstation.

Models

Path	Size	Description
svs/model.safetensors	~2.82 GB	Singing Voice Synthesis (lyrics+MIDI → singing)
svc/model.safetensors	~2.79 GB	Singing Voice Conversion (audio-to-audio)
whisper-base/	~0.29 GB	Verbatim openai/whisper-base (Apache-2.0) — the SVC model's frozen semantic encoder. Its weights are not in the SVC checkpoint; bundling them here keeps SVC fully offline.
config.yaml	579 B	Model architecture configuration
phone_set.json	~30 KB	Phoneme mapping for SVS

Preprocessing assets

The SVS annotation pipeline (vocal separation → RMVPE F0 → lyric ASR → ROSVOT note transcription) pulls its component weights on demand from the companion repo AEmotionStudio/soulx-singer-preprocess (verbatim upstream files from the SoulX-Singer-Preprocess release).

Architecture

Flow-matching based (F5-TTS foundation)
22-layer transformer with 1024 hidden size, 16 heads
128-dim mel spectrogram, 24kHz output
Trained on 42,000+ hours of aligned vocals (Mandarin, English, Cantonese)

Conversion notes

svs/ and svc/ are torch.load(ckpt)["state_dict"] → safetensors conversions of upstream model.pt / model-svc.pt; loaders consume them with strict=True.
whisper-base/ is byte-identical to upstream openai/whisper-base (sha256-verified at upload; see backend/scripts/mirror_soulx_singer_to_hf.py in MAESTRO).

License

Apache 2.0