A 230 MB Vocaloid Voice on Apache-2.0: What the Licence Actually Covers

What happened

NeatAvocado14 published KasaneTeto-GPTSoVITS-V2 to the Hugging Face model hub on 23 August 2026. It is a GPT-SoVITS V2 fine-tune trained on Japanese audio of the Vocaloid character Kasane Teto, totalling roughly 230 MB across two weight files.

Context

GPT-SoVITS has become the default open-source TTS stack for voice cloning and character-specific synthesis. The V2 architecture splits the job into two stages: a GPT-based language model generates semantic tokens, then a SoVITS vocoder renders the waveform. Fine-tunes on specific voices have been appearing steadily since the base repository opened on GitHub, but most target real-person voices or generic multilingual models. A fine-tune locked to a single Vocaloid character, trained exclusively from bilibili-sourced Japanese audio, is a different category: it is a character asset as much as a speech model, and the rights question is different from a voice clone of aYouTuber or a podcast host.

How it works

The model ships as two files: a 155 MB GPT checkpoint (bilibili_jp_20260823-e20.ckpt) and a 75 MB SoVITS vocoder (bilibili_jp_20260823_e4_s496_l32.pth). Inference is reference-audio-based. You supply a reference clip (the bundled ref_audio.wav) plus its Japanese transcript, and the pipeline conditions on that to synthesise new text in the same timbre. There is no standalone binary. You clone the RVC-Boss/GPT-SoVITS repository from GitHub, drop each weight file into its expected directory, and run inference_webui.py or the v2 inference UI. A sample output, preview_demo.wav, is included as a synthesis preview. No VRAM floor, no real-time factor, no Docker image, and no REST API are documented in any source. The model is locked to this one trained voice; the card does not describe a multi-speaker or voice-cloning workflow beyond the reference-audio conditioning it already uses.

Our read

The Apache-2.0 tag on the Hugging Face page and the model card's own disclaimer are in direct tension, and that tension is the whole story. The card states the project is "for learning and entertainment purposes only" and assigns the voice and timbre to the character's original rights holder, with users assuming responsibility for their use. Kasane Teto is a Vocaloid character with a defined rights holder. The Apache-2.0 licence governs the weight files as code artefacts; it does not clear the underlying voice data or the character IP. A small studio that loads these weights into a client deliverable is taking on a liability the licence text does not cover, and the model card's own language is an admission of that gap.

The second-order effect is more interesting. At zero downloads and zero likes, this is a single user's upload with no community signal. But the pattern matters: character-specific TTS fine-tunes built on scraped platform audio are going to multiply, and the question the industry has not yet answered is whether a permissive code licence on the weights is sufficient to make a character voice commercially synthesisable. The model card says no. A court in most jurisdictions would likely agree, because the licence on the weights and the rights in the training data are separate legal objects.

For a studio running ComfyUI: there is no custom node. You need the Python WebUI or an API wrapper running as a sidecar service. That is a pipeline integration, not a drop-in, and it adds a process to babysit on every render.

What this changes

Nothing in a shipping pipeline changes on Monday. If you need a Japanese voiceover for a client video, this is not the asset you reach for: the character-IP question is unresolved, the bilibili audio provenance is unverified, and there is zero community validation at 0 downloads and 0 likes. For internal prototyping — benchmarking the GPT-SoVITS V2 quality floor for Japanese synthesis, or testing reference-audio conditioning against your existing TTS stack — 230 MB of weights is cheap to load alongside a video model on the same GPU, provided VRAM allows. The tradeoff is integration effort: no ComfyUI node, no Docker, just the Python WebUI and a directory structure you have to maintain.

License

Apache-2.0, as stated on the Hugging Face model card. Commercial use of the weight files as code artefacts is permitted. The model card separately restricts use to non-commercial, non-infringing scenarios and assigns the voice and character to the original rights holder, which overrides the permissive weight licence for any commercial deliverable involving this specific voice.

Key takeaways

• A 230 MB, two-stage GPT-SoVITS V2 fine-tune on a single Vocaloid character, published 23 August 2026 with zero community engagement at the time of collection.
• The Apache-2.0 weight licence does not clear the character IP or the bilibili-sourced training data; the model card restricts use to learning and entertainment and assigns the voice to the character's rights holder.
• Inference requires the GPT-SoVITS Python repository and its WebUI; no standalone binary, Docker image, REST API, or ComfyUI node is documented.
• For a small studio: not a client-facing asset. A lightweight benchmark target for Japanese TTS quality and reference-audio conditioning if the integration cost is acceptable.
• No VRAM requirement, inference speed, or quality comparison to production Japanese TTS engines is stated in any source.

Sources

1. NeatAvocado14/KasaneTeto-GPTSoVITS-V2 — text-to-speech on Hugging Face — tier 3