neutts-jp-150m-8bit

NeuTTS-JP-150M — INT8 ONNX weights

This Hugging Face model repository hosts INT8-quantized ONNX weights for NeuTTS-JP-150M, a compact Japanese text-to-speech stack that supports reference-based voice cloning when paired with a NeuCodec waveform codec and the same tokenizer assets as the FP16 / FP32 ONNX bundles.

It is published as a separate repo from aoiandroid/neutts-jp-150m-onnx-float16 and aoiandroid/neutts-jp-150m-onnx-float32 so clients can download a smaller, int8-focused bundle into a dedicated on-disk directory.

Contents (expected layout)

Typical files on main:

NeuCodec encoder / decoder Core ML or ONNX packages are not required to live in this repo; consuming apps often fetch them from their own NeuCodec distribution. This repo is scoped to the JP-150M ONNX LM + tokenizer sidecar consistent with other JP-150M ONNX variants.

Downstream integration (on-device iOS client)

The iOS / macOS app on-device iOS client resolves ONNX in this order for the INT8 Hub variant:

1. model_int8.onnx
2. model.onnx
3. neutts-jp-150m-8bit.onnx
4. neutts-jp-150m-original.onnx
5. model_fp16.onnx (fallback if published)

Repo ID in code: aoiandroid/neutts-jp-150m-8bit
On-disk folder label: NeuTTS_JP_150M_ONNX_INT8

Related artifacts

Usage notes

• Language: Japanese (ja).
• Reference audio: Short clean speech (consumer apps typically require ≥ ~1 s of reference audio for stable cloning).
• Runtime: ONNX Runtime with operators matching the export; INT8 graphs may use reduced dynamic range vs FP16 — validate quality on your target devices.

Citation / upstream

If you use NeuTTS-JP-150M in research or products, cite the upstream NeuTTS / Neuphonic materials (model card, paper, or official repository) that match the base checkpoint you quantized. This repo only redistributes converted ONNX artifacts for app delivery.

License

See the LICENSE file in this repository (and any upstream license terms for the base NeuTTS-JP-150M checkpoint). The YAML license: other is a placeholder until the SPDX identifier is aligned with that file.