SoulX-Singer GGUF
Local singing synthesis (SVS), singing voice conversion (SVC), and ROSVOT singing-note preprocessing for audio.cpp. SVS and SVC produce 24 kHz mono audio and support Mandarin, English and Cantonese. Inference is offline, not real-time streaming.
Upstream weights: Soul-AILab/SoulX-Singer,
revision 40493ad90286056c7a9095035164434a79daa8c9.
Reference implementation: Soul-AILab/SoulX-Singer,
commit 81aeb3ae772c70093c3de74dc23c92d983801ae4.
Packages
| Task | F16 | Mixed Q8_0 |
|---|---|---|
| Singing synthesis | soulx-singer-f16.gguf | soulx-singer-q8_0.gguf |
| Singing voice conversion | soulx-singer-svc-f16.gguf | soulx-singer-svc-q8_0.gguf |
Quantization applies to DiT matrices. Vocos matrices remain FP16; normalization, conditioning and convolution weights remain FP32. SVC also bundles the FP32 Whisper-base content encoder and RMVPE pitch extractor. ROSVOT is supplied separately in ROSVOT-GGUF.
Run
Both Singer packages use the soulx_singer family. Select the package matching
the task. A build with SOXR is required for SVC pitch preprocessing.
Singing voice conversion
Use isolated source vocals and reference audio for the target voice. Accompaniment separation and mixing are separate steps.
audiocpp_cli --task svc --family soulx_singer \
--model SoulX-Singer-SVC-GGUF/soulx-singer-svc-q8_0.gguf \
--backend cuda --threads 8 \
--audio singing.wav --voice-ref reference.wav \
--request-option seed=2026 --out converted.wav --log
Singing synthesis
Provide reference singing plus matching reference metadata and target phoneme/note metadata in the upstream format. Plain lyrics or MIDI alone are not accepted. Metadata preparation is separate from synthesis.
audiocpp_cli --task gen --family soulx_singer \
--model SoulX-Singer-GGUF/soulx-singer-q8_0.gguf \
--backend cuda --threads 8 --voice-ref reference.wav \
--request-option reference_metadata_file=reference.json \
--request-option target_metadata_file=target.json \
--request-option control=score --request-option seed=2026 \
--out singing.wav --log
Defaults are 32 Euler steps and guidance 3. SVS defaults to control=melody;
control=score follows note pitches. Automatic pitch matching is off by default;
use auto_pitch_shift=true to enable it. Other request options include
num_steps, guidance_scale, pitch_shift and seed.
Use --backend vulkan or --backend cpu to select another validated backend.
CUDA performance
RTX 5090, eight threads, Debug build, warmed requests, 32 steps and guidance 3. These comparisons explicitly enable automatic pitch matching; SVS uses score control. Python uses official FP16 inference. Timing includes conditioning, denoising and vocoding, excluding model loading and WAV serialization. Peak VRAM covers the whole process, sampled every 20 ms. Lower RTF is faster; RTF 0.1 means one second of computation per ten seconds of audio.
| Workload | Implementation | RTF | Peak VRAM |
|---|---|---|---|
| SVC, 51.48 s | Python FP16 | 0.08508 | 4,416 MiB |
| SVC, 51.48 s | audio.cpp F16 | 0.07415 | 4,532 MiB |
| SVC, 51.48 s | audio.cpp mixed Q8_0 | 0.06649 | 4,150 MiB |
| SVS score, approximately 8 s | Python FP16 | 0.10070 | 3,492 MiB |
| SVS score, approximately 8 s | audio.cpp F16 | 0.06909 | 2,342 MiB |
| SVS score, approximately 8 s | audio.cpp mixed Q8_0 | 0.06656 | 1,982 MiB |
Mixed Q8_0 reduces generation time by approximately 22% for SVC and 34% for SVS versus these Python runs. SVS peak VRAM is approximately 43% lower.
CPU and Vulkan results
Both tasks completed with mixed Q8_0 packages and produced the expected Mandarin lyrics in ASR checks. These are single eight-second default-setting validation requests, including first-use graph setup, rather than warmed benchmarks. They use no automatic pitch matching; SVS uses its default melody control. Their settings and clip lengths differ from the CUDA comparison above.
| Backend | Task | Output duration | Request time | RTF | Output check |
|---|---|---|---|---|---|
| Vulkan, RTX 5090 | SVC | 8.00 s | 1.05 s | 0.13131 | Expected lyrics recognized |
| Vulkan, RTX 5090 | SVS melody | 8.00 s | 0.59 s | 0.07422 | Expected lyrics recognized |
| CPU, Ryzen 9 7900X | SVC | 8.00 s | 71.43 s | 8.92874 | Expected lyrics recognized |
| CPU, Ryzen 9 7900X | SVS melody | 8.00 s | 55.55 s | 6.94399 | Expected lyrics recognized |
Audio examples
Normal audio.cpp mixed-Q8 outputs, seed 2026, 32 steps and guidance 3. The SVC conversion and Mandarin SVS score example enable automatic pitch matching. SVC uses the source vocals and reference voice below; SVS uses aligned reference/target metadata. The English and Cantonese examples use upstream example metadata. Outputs are generated vocals without added accompaniment.
Reference voice 路 Download WAV
Source singing for SVC 路 Download WAV
SVC converted vocals, mixed Q8_0 路 Download WAV
SVS Mandarin, mixed Q8_0 路 Download WAV
SVS English, mixed Q8_0 路 Download WAV
SVS Cantonese, mixed Q8_0 路 Download WAV
ROSVOT preprocessing
Offline singing-note transcription for audio.cpp. One F32 package contains ROSVOT, RWBD word-boundary prediction and RMVPE pitch extraction. Outputs are MIDI or JSON note events; lyric transcription and streaming are not supported.
Checkpoint source:
Soul-AILab/SoulX-Singer-Preprocess,
revision 83dc50289d22a81b1e9998f5b9e111aef7c1fdcd.
Reference implementation: ROSVOT,
commit 3c8332bf43adae35f6e4d64971862f2f6139b310.
The source model card declares Apache-2.0 for the checkpoint distribution;
ROSVOT reference code is MIT licensed. License notices are embedded in the GGUF.
audiocpp_cli --task midi --family rosvot \
--model ROSVOT-GGUF/rosvot-f32.gguf \
--backend cuda --threads 8 --audio singing.wav --out notes.mid --log
Use --request-option output_format=json --out notes.json for note pitches,
start/end seconds and word indices. An optional word_durations_file JSON array
provides word timing instead of predicted RWBD boundaries. Use isolated vocals.
Validated on CPU, CUDA and Vulkan. CPU component parity passes the strict FP32 comparison; GPU rounding can change discrete note boundaries. This is a note-transcription model, not a guarantee of perfect musical transcription.
License
SoulX-Singer is Apache-2.0 licensed. Package directories include upstream and dependency notices, also embedded in the GGUFs. ROSVOT attribution is described in its preprocessing section above.
- Downloads last month
- 7
8-bit
16-bit
32-bit
Model tree for audio-cpp/SoulX-Singer-GGUF
Base model
Soul-AILab/SoulX-Singer