Vocalist

Vocalist is a native macOS app for voice cloning and conversion. No Python, no containers, no server: the entire stack, training included, runs on-device on Apple Silicon through Apple's MLX framework.

It is in private development for now. If voice conversion on Mac is your thing, keep an eye on the project: github.com/froggeric/vocalist

What this repository is

Pretrained base models are the fuel of voice cloning, and several of the best ones sit on links that can vanish without notice. This is a curated, checksummed mirror of the bases Vocalist trains from, so they stay available whatever happens upstream. All credit for the training quality belongs to the original creators, named per model below; this mirror only keeps the lights on.

Each pair is verified before upload: the archive structure is inspected and recorded in manifest.json (sample rate, vocoder architecture, iteration, dtype, SHA-256 for every file).

Pretrained pairs

Base	Vocoder	Rate	Source (original release)
refinegan_vctk_v1	RefineGAN, original RVC variant	44.1 kHz	SimplCup/RefineGanVCTKV1 (VCTK, 35 h, 177 epochs)
refinegan_klm50_exp1	RefineGAN, original RVC variant	44.1 kHz	SeoulStreamingStation KLM 50 exp1 (300 h; recovered via the Politrees/RVC_resources git history)
hifigan_klm49	HiFi-GAN NSF	48 kHz	SeoulStreamingStation KLM49_HFG (provenance unconfirmed)
G/D_KLM_RVC_PT_48k	HiFi-GAN NSF	48 kHz	SeoulStreamingStation KLM-HF Trainer (the author's fine-tune of the official 48k base, Korean music-cover/vocal corpus; hosted as fp16)
hifigan_titan_medium	HiFi-GAN NSF	48 kHz	community release (provenance unconfirmed)
ik_spanish_f0	HiFi-GAN NSF	48 kHz	IK Multimedia reSing in-house model (extracted from the product; reSing is a UI over stock RVC). Eval-only; proprietary origin

Files live under pretrained/ as <name>_G_<rate>.pth / <name>_D_<rate>.pth pairs (generator and discriminator).

A note on the 44.1 kHz pairs

Both 44.1 kHz bases use the original RVC RefineGAN architecture: harmonic sine source with 9 components, ResBlock down-ladder, biased output convolution. Vocalist supports fine-tuning from these out of the box, including the empirical 100 fps geometry (hop 441) the checkpoints were trained with. The vctk pair is a clean 177-epoch release; the klm50 archive carries optimizer state from a mid-training resume (its iteration field is stale), which fine-tuning ignores.

A piece of history this mirror preserves: the KLM 50 experimental 44.1 kHz line (exp2, exp3, and the final XE4 of January 2025) lived only in a HuggingFace repository its author later deleted. Those versions are, as far as we can determine, unrecoverable. The exp1 pair here was recovered from the git history of a mirror repository before it, too, could disappear. The author's maintained KLM RefineGAN line continued at 32 and 40 kHz.

Using the models elsewhere

These are standard RVC training checkpoints in torch format. Any RVC-lineage tool that accepts .pth pretrained pairs can use them; the manifest records everything a loader needs to know about each archive.

Interest?

Vocalist trains and converts voice models entirely on Mac hardware, at up to 44.1 kHz. If you would like early access or just want to watch it come together, star the GitHub repository or follow @froggeric.