SBV2 Core ML shared BERT and dictionary
Use this shared Japanese BERT/dictionary with the SDK and samples and a separate voice model.
Choose a variant
int8/ contains INT8; float32/ contains original FP32. Each folder includes BERT, the dictionary, licenses and its download manifest.
Both variants use the same compatible voice models.
| Variant and files | Shared size | With JVNV | Choose when | App manifest |
|---|---|---|---|---|
| INT8 | About 503 MB | About 797 MB | Storage size matters | Download INT8 |
| Original FP32 | About 1.52 GB | About 1.81 GB | Avoiding weight quantization or shorter preparation matters | Download FP32 |
INT8 stores BERT weights in 8bit. BERT computation and tensor interfaces are FP32 for both variants. Requires iOS 18+ or macOS 15+ on Apple Silicon. Compilation caches need additional storage; file savings do not establish RAM savings.
Download in the sample app
Open モデル設定 → URLからモデルを取得 → 共通モデル and choose INT8 or FP32 under 共通モデルの版. The URL is filled automatically; press ダウンロード. For a custom URL or SDK download, use these pinned manifests:
INT8
https://huggingface.co/AILogDev/sbv2-coreml-common/resolve/973d6e239af305f0d78a8bf30c6af5093c0fd47d/int8/download.json
Original FP32
https://huggingface.co/AILogDev/sbv2-coreml-common/resolve/973d6e239af305f0d78a8bf30c6af5093c0fd47d/float32/download.json
Choose a separate voice and press モデルを準備 after download. Synthesis works offline.
Download one variant on Mac
Use the Hugging Face CLI.
Keep --include: downloading the entire repository includes both variants and compatibility files.
INT8 → models/int8/
hf download AILogDev/sbv2-coreml-common \
--revision 973d6e239af305f0d78a8bf30c6af5093c0fd47d \
--include "int8/*" --local-dir models
Original FP32 → models/float32/
hf download AILogDev/sbv2-coreml-common \
--revision 973d6e239af305f0d78a8bf30c6af5093c0fd47d \
--include "float32/*" --local-dir models
Select the downloaded int8/ or float32/ root in the sample app. For the SDK, pass its bert/ and dictionary/ subfolders and a separate voice folder.
Download complete .mlpackage contents.
Switch models and compare
Stop synthesis/playback, choose the other shared root in 共通モデル, and press モデルを準備.
The shared-folder label shows INT8/FP32 from its metadata. Choosing a download variant alone does not switch a prepared model.
For the SDK, await completion of existing work, then call load and warmUp with the selected shared folder.
Old folders and caches are not removed automatically. Quit the app or call unload before deleting an unused shared folder; retain your separate voice folder.
| iPhone 17 Pro / JVNV Neutral example | INT8 | Original FP32 |
|---|---|---|
| Cached reload + warm-up | About 18 seconds | About 6 seconds |
| Warmed median RTF, three sentences | 0.071 | 0.067 |
Initial compilation, playback and concurrent LLM execution are excluded. Texts, voices and devices vary; INT8 is not necessarily faster. BERT rounding can change prosody, timing, duration and waveform. Voice weights and the dictionary are unchanged. No concerning difference was heard in the JVNV Neutral comparison clips; compare your intended voices and texts.
SDK guide · Getting started · Create compressed shared models
Previous URLs
Previous pinned URLs remain usable. Root download.json, bert/ and dictionary/ retain INT8 compatibility for existing URLs.
For new downloads, choose int8/ or float32/ above.
Attribution and terms
BERT derives from Kyoto University's NLP group's Japanese DeBERTa, revision 547b0e8b044fba3f9b84d0ab9f990440bd130c8b.
BERT is CC BY-SA 4.0; see each variant's LICENSE.md and provenance.json.
The Open JTalk 1.11 UTF-8 dictionary has separate BSD-style terms in each variant's dictionary/COPYING.
File hashes are in checksums.json. Code is AGPL-3.0. This unofficial conversion is not endorsed by the original authors.