New features, model additions and replacements, performance improvements and fixes, recorded as they ship.
Choose a model family, then its version or processing mode. The demo follows your selection.
Related separation and MIDI versions use two levels of cards, with individual status and settings. MuScriptor defaults to segmented repair; YourMT3+ to YPTF.MoE+Multi noPS. Multi-instrument MIDI demos use the full song だから僕は音楽を辞めた. MIDI previews also load correctly when filenames contain special characters.
Upload recordings, check your material and review the credit quote to train a private voice for song covers, or fine-tune a compatible existing model.
Open RVC model training directly from Voice and Singing. The dedicated page shares material checks, quotes, training progress, history retries and model saving with the cover page, and keeps your existing training draft.
Prevent high audio peaks from causing vocal and other stem exports to fail.
Chorus male/female vocal separation checks peaks across the complete stem set before export. When attenuation is needed, every stem receives the same constant gain, preserving relative levels, dynamics and full duration. Results that need no attenuation keep their original level. The shared export fix also covers crowd removal, denoising, echo removal and DrumSep drum separation. The downloadable audio-export.json records the actual gain adjustment.
Training starts when ordinary tasks have finished and resources are available. New tasks pause training after saving a checkpoint; it resumes when resources are free. Pausing and resuming add no credits.
Move the pointer, touch the chart or drag the slider to inspect values. The slider supports arrow keys, Home and End. Continuing the same task restores the saved model, optimizer and training position. After an unexpected interruption, steps since the last confirmed save may be repeated. Upload an existing RVC model here. If you only have recordings, use “Train my voice”. Uploads are private by default; other users can select a model only after you enable public sharing.
Select separation and MIDI models individually and see their status, parameters and published results.
IDM, MSG-LD, ADTOF, ADT_STR and other models have individual cards, with task names reflecting the selected model. Each category chart includes every model and compares scores only under matching test conditions; missing results are marked explicitly. OaFS reports completed feature windows, inference batches and decoding work. Audio-track and MIDI playback gain now reaches +6 dB, with improved controls on phones. Added demos using the same complete input for comparable models and regenerated the AI cover demo, preserving lossless audio. ADTOF Direct accepts isolated drums or a full mix, extracting drums first for full mixes. Fixed voice lists staying empty after a temporary refresh failure.
More MIDI transcription choices for full mixes and piano recordings.
OaFS detects 34 instrument classes and exports MIDI and note tables with fixed velocity 127. SFT-CRNN detects piano note onsets and offsets, with fixed velocity 100 and no pedal prediction.
Separate drums with IDM; transcribe hits with ADT_STR and MSG-LD.
IDM exports nine drum stems without MIDI. ADT_STR detects onsets and velocities for 26 drum classes and exports MIDI. MSG-LD generates five drum stems and detects onsets; MIDI uses fixed velocity 100 and playback duration 0.1 seconds.
HookKern adds quartets; Tsumugi and SwiftF0 are updated.
HookKern keeps its melody-and-chord model and adds string quartets. Tsumugi adds vocal harmony v1.5/v1.6 and drums v1.5, with harmony v1.6 as the default. SwiftF0 moves from 0.1.2 to 0.2.0.
Preserve short notes, audition new drum stems and explore real demos.
Very short MIDI notes now end correctly; notes whose timing was not edited retain their original timing on export. Preview nine IDM stems or five MSG-LD stems on the results page. OaFS, SFT-CRNN and IDM demos now use real task outputs.
Material checks show usable duration, and interrupted training retains progress.
Checks show usable audio, the required duration and how much more to add. Failed tasks can be retried. Training queues and checkpoints survive service restarts and resume when scheduling conditions are met.
Audio decoding fixes and separate downloads for MIDI already processed.
Audio decoding has been updated. If a full transcription does not finish, the task page offers generated segments clearly marked as incomplete. The failed task keeps its refund status.
Single-instrument MIDI, melody-and-chord scores and voice synthesis are available again.
Basic Pitch turns single-instrument audio into MIDI. HookKern produces melody-and-chord scores, and ZONOS2 synthesizes speech from a reference voice. Their demos now use outputs from real tasks in this release.