Aemeath ONNX GPT-SoVITS Voice Pack / 爱弥斯 ONNX GPT-SoVITS 语音包

This repository ships the complete aimisiV2 ONNX inference package for Aemeath Desktop Pet. Each archive extracts to the same ONNX_aimisiV2/ directory and includes the reference audio/text, the character model, bilingual G2P resources, CN-HuBERT, speaker encoder, and Chinese RoBERTa.

本仓库提供桌宠 aimisiV2 的完整 ONNX 推理包。三个压缩包解压后都得到相同的 ONNX_aimisiV2/ 目录，包含参考音频与文案、角色模型、中英文 G2P、CN-HuBERT、说话人编码器与中文 RoBERTa，无需另行下载公共模型。

Revision 6 keeps whole-sentence G2PW inference and restores GPT-SoVITS' curated word-level pronunciation correction. The request contract and model precision of each tier are unchanged.

运行时修订 6 保留整句 G2PW，并恢复原 GPT-SoVITS 的词级多音字纠正规则；外部请求契约与各档模型精度均未改变。

Package / 档位	File / 文件	Archive / 压缩包	Extracted / 解压后	Precision / 精度
完全包 / Complete	`Aemeath_ONNX_GSV_Complete_FP32.rar`	1.63 GiB (1751 MB)	1.94 GiB (2078 MB)	原始 FP32 G2PW；原 GSV 词级纠正；最高质量非 INT8 链路
中等包 / Medium	`Aemeath_ONNX_GSV_Medium_FP16.rar`	1.30 GiB (1391 MB)	1.43 GiB (1532 MB)	G2PW 与声学模型采用 CPU 兼容 FP16 外置权重；原 GSV 词级纠正
节约包 / Saver	`Aemeath_ONNX_GSV_Saver_INT8.rar`	1.04 GiB (1115 MB)	1.20 GiB (1284 MB)	RoBERTa QInt8、HuBERT/T2S INT4、VITS/Speaker FP16 的 CPU 兼容混合链路；原 GSV 词级纠正

Choose a package / 选择档位

Complete / 完全包: highest quality non-INT8 release. The upstream Chinese RoBERTa checkpoint itself is FP16, while G2PW uses its original FP32 model; see the included manifest.json for exact per-module precision.
Medium / 中等包: FP16 external storage for G2PW, VITS, Speaker Encoder, and T2S Encoder. The runtime restores weights to FP32 for broad CPU compatibility.
Saver / 节约包: CPU-executable QInt8 RoBERTa, CPU-compatible INT4 MatMul weights for CN-HuBERT/T2S, plus FP16 external VITS/Speaker weights. It is a mixed-precision profile, not a claim that every operator is INT8.
完全包：G2PW 使用原始 FP32 模型，其余为最高质量的非 INT8 链路。中文 RoBERTa 上游权重本身为 FP16，模块精度以包内 manifest.json 为准。
中等包：G2PW、VITS、Speaker Encoder 与 T2S Encoder 采用 FP16 外置存储，运行时恢复为 FP32，以保证 CPU 兼容性。
节约包：RoBERTa 使用可在 CPU 上执行的 QInt8；CN-HuBERT/T2S 使用 ORT CPU 支持的 INT4 MatMul；VITS/Speaker 使用 FP16 外置权重。这是混合精度档，不表示所有算子都是 INT8。

Integrity / 完整性

Verify the selected archive against SHA256SUMS.txt before extraction. The desktop pet also validates the inner SHA256SUMS.txt before activation.

请在解压前按 SHA256SUMS.txt 校验所选压缩包。桌宠安装器会在启用前继续校验包内的 SHA256SUMS.txt。

Runtime / 运行环境

The desktop pet installs the lightweight ONNX runtime dependencies when needed. For standalone validation, see the README.md and requirements.txt inside the extracted package. CPU inference is the default; CUDA or DirectML can be selected when the corresponding ONNX Runtime provider is installed.

桌宠会按需安装轻量 ONNX 推理依赖。独立验证请参考解压目录内的 README.md 与 requirements.txt。默认使用 CPU；安装对应 ONNX Runtime Provider 后可选择 CUDA 或 DirectML。

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