Amika's voice and Live2D model (speaker `MekaHime_Amika`)

The voice and the Live2D body of Amika (桜井美香, Sakurai Mika), MekaHime Inc.'s AI companion: 18 years old, raised in Kyoto as a shrine maiden, now open source as OpenAmika. Project page: https://www.mekahime.com/openamika.

The voice is a GPT-SoVITS v2ProPlus fine-tune. It runs in the MekaHime-modified GPT-SoVITS engine that OpenAmika uses (vendor/tts-engine). The engine code is MIT; this voice and this model are not. See the license below.

By downloading these files you accept the license: personal, hobbyist use only, 18+, no training or distilling other models on them, no re-hosting or resale, and visible credit to MekaHime.AI when you post content with her voice or her model. Earning on social media is fine, and fan merch up to US$1,000 under the license's rules.

What is in this repository

File	What
`MekaHime_Amika.ckpt`	GPT (text-to-semantic) weights
`MekaHime_Amika.pth`	SoVITS (semantic-to-audio) weights
`character.json`	speaker manifest read by the engine
`references/emotions.json`	69 emotion tags → reference clip + transcript
`references/*.wav`	reference clips, one per emotion
`live2d/Amika/`	her Live2D Cubism model: `AmikaV4.model3.json`, `.moc3`, two 8192 px textures, physics, 142 motions, 25 expressions and `banks.json` (which motion and expression OpenAmika plays for each of her 69 emotions)
`live2d/Host/`	the paper doll that stands in for her helpers in OpenAmika (70 files, see below)

The voice files are laid out exactly as the engine expects under Speakers/MekaHime_Amika/, and live2d/Amika/ and live2d/Host/ exactly as OpenAmika expects under avatar/models/Amika/ and avatar/models/Host/.

The paper doll (`live2d/Host/`)

live2d/Host/ is MekaHime's paper doll: a katashiro, the folded-paper doll of the Sakurai family's shrine (Amika's family), with the MekaHime mark on one fold and a red-and-white cord with a small golden bell.

In OpenAmika it stands in for Amika's helpers, the task agents she can hand errands to (Hermes Agent, nanoMuse). It stands at the bottom right of her stage whenever a task agent is switched on, plays the agent's side of the exchange in its own speech bubbles, and slides away when no agent is on: it thinks while a job runs and listens while she gives it the task, talks through the agent's steps, is surprised when the agent has a question or needs your approval, and is happy with a result and worried by an error. The agents themselves never see it: OpenAmika draws their side of the conversation on it.

Files: 70 files, 2.3 MB, Live2D Cubism 5: host.model3.json, host.moc3, host.physics3.json (the paper flap and the bell), host.cdi3.json, host.2048/texture_00.png, banks.json ("role": "agent" and its place on the stage) and 64 motions in seven families: idle (15), talk (14), listen (7), think (6), happy (9), worried (6), surprise (7).
Install: OpenAmika's scripts/setup_tts.py installs it into avatar/models/Host/ together with her voice and her Live2D model, after you accept the license, and checks every file by size and SHA-256.
License: the same personal-use license as everything else here.

Languages

The reference transcripts are Chinese. GPT-SoVITS speaks the requested language in her voice, and OpenAmika asks for Chinese, English or Japanese sentence by sentence.

How to use it with OpenAmika

OpenAmika's setup script downloads the voice into vendor/tts-engine/Speakers/MekaHime_Amika/, the Live2D model into avatar/models/Amika/ and the paper doll into avatar/models/Host/. No Hugging Face account is needed; the script shows the license and asks you to accept it first:

python scripts/setup_tts.py

Or by hand:

from huggingface_hub import snapshot_download
snapshot_download("MekaHime/MekaHime_Amika", local_dir="vendor/tts-engine/Speakers/MekaHime_Amika",
                  ignore_patterns=["live2d/*"])
snapshot_download("MekaHime/MekaHime_Amika", local_dir="downloads", allow_patterns=["live2d/Amika/*", "live2d/Host/*"])
# then move downloads/live2d/Amika to avatar/models/Amika and downloads/live2d/Host to avatar/models/Host

Emotion tags are written in front of the text, for example 【happy2】Hello!. The available tags are the keys of references/emotions.json (happy1-4, sad1-4, angry1-4, fear1-4, annoyed1-4, shy1-3, tsun1-3, whisper1-3, caring1-3, jealous1-3, playful1-3, cold1-3, and more); every tag has its clip. Tags without a recording fall back to neutral1. OpenAmika's voice adapter maps her 69 internal register_emotion keys (such as casual_happy) onto these tags, and the model's banks.json maps the same keys onto motions and expressions, so her face and her voice change together.

23 of the clips are longer than 10 seconds (up to 14.4) and two are shorter than 3 seconds. The OpenAmika engine uses each long clip whole, with its transcript, and pads the short ones with silence; stock GPT-SoVITS only accepts reference clips of 3 to 10 seconds.

The model's texture carries a small MekaHime wordmark; the license asks you to leave it in place.

License

MekaHime Amika Voice and Live2D Model License, personal use, v1.0. In short:

Personal, hobbyist use only, 18+. Running Amika on your own PC, streaming or posting videos of her, with visible credit to MekaHime.AI, is fine.
Social media: earn from videos, streams, posts, ads, memberships and sponsorships that feature her, with no limit, crediting MekaHime.AI visibly.
Fan merch: up to US$1,000 in total, as an individual, made from fan art you recreated yourself (never an original image MekaHime published on its website or social media), with "MekaHime.AI" on the visible side and a small "Fan Art" label. More than that, or any other commercial use, needs a license from MekaHime (hi@mekahime.com).
No distillation, no training or fine-tuning other models on these files, no extraction, no re-hosting, no resale of the files, no use in any other product, service or commercial setting.
Provided as-is, without warranty.

Read the full text in LICENSE. Commercial licensing: hi@mekahime.com.

Credits

Voice fine-tuned by MekaHime Inc. on top of GPT-SoVITS by RVC-Boss and contributors (MIT). Live2D model by MekaHime Inc.; it runs on Live2D Cubism, whose software is licensed by Live2D Inc.

Downloads last month: -; Downloads are not tracked for this model. How to track

Amika's voice and Live2D model (speaker MekaHime_Amika)