# Shared BERT and dictionary for SBV2 Core ML

[日本語](README.md)

For [SBV2 Core ML](https://github.com/Corvelis/sbv2-coreml) on iOS 18+ and macOS 15+ Apple Silicon.
Download the [companion package](https://huggingface.co/AILogDev/sbv2-coreml-jvnv-f1-jp) as well.
Package size is approximately 1.52 GB; allow extra space for first-use Core ML compilation.

## Download and use

In the sample's **モデル設定 → URLからモデルを取得**, choose the package type and enter:

```text
https://huggingface.co/AILogDev/sbv2-coreml-common/resolve/main/download.json
```

Alternatively use the [Hugging Face CLI](https://huggingface.co/docs/huggingface_hub/guides/cli):

```sh
hf download AILogDev/sbv2-coreml-common --local-dir models/sbv2-coreml-common
```

Download the entire folder, including the contents of `.mlpackage` directories.
Pass `bert/` and `dictionary/` to `ModelPaths`. Select the package root as **共通モデル** in the sample.
Speech synthesis works offline after downloading both packages.

[Quick Start](https://github.com/Corvelis/sbv2-coreml/blob/main/docs/getting-started.ja.md) ·
[SDK integration](https://github.com/Corvelis/sbv2-coreml/blob/main/docs/sdk-guide.ja.md)

## Attribution and licenses

BERT by Kyoto University NLP group: [original model](https://huggingface.co/ku-nlp/deberta-v2-large-japanese-char-wwm), revision `547b0e8b044fba3f9b84d0ab9f990440bd130c8b`. Converted to two FP32 Core ML models with input lengths 64/128/256, without retraining or quantization. BERT is CC BY-SA 4.0. The Open JTalk 1.11 UTF-8 dictionary has separate BSD notices in `dictionary/COPYING`.

See [LICENSE.md](LICENSE.md), `provenance.json` and `checksums.json` for licenses,
conversion details and SHA-256 checksums. This is an unofficial conversion and is not
endorsed by the original creators. SDK, sample and converter code is **AGPL-3.0**.