v3.2.1

v3.2.1

• Faster and more stable QNN T2S decoding with double-buffered KV cache.
• VITS partition outputs are passed directly between QNN stages to reduce tensor copies.
• Verified on Snapdragon 8 Elite; typical end-to-end synthesis is about 7 seconds.
• Release APK is unsigned for local installation/testing.

QNN remains an engineering preview and requires matching pipeline/voice attachments.

GSV Mobile v3.2.0

GPT-SoVITS 最好的手机优化版本。

• V2 Pro Plus 最低 8 GB 物理内存（已验证）
• 新增真实设备截图与性能对比
• Android CPU FP32 与 QNN HTP 工程支持

v3.1.0

• Rebuilt the mobile UI around synthesis, model setup, and settings workflows.\n- Added a recommended Firefly V2 Pro Plus setup path. It downloads the CPU fallback plus only the matching Snapdragon QNN attachments for the detected device, verifies SHA-256, and stores all artifacts in /sdcard/models/gs.\n- Added clear pipeline selection and model/QNN attachment management.\n- QNN remains a preview while broader device validation continues.

GSV Mobile v3.0.0

v3.0.0

• Chinese and English UI with automatic system-language selection and an in-app override.
• Runtime-scoped reference audio in the app and local OpenAI-compatible API; the converted reference remains the default.
• Model discovery is restricted to /sdcard/models/gs.
• Updated pipeline/model package validation and download progress.

QNN/NPU status

QNN/NPU artifacts are included only for engineering preview and hardware validation. They are not a supported v3 feature: matching physical-device inference, profiling and audio acceptance remain pending. Use the CPU .gsvm packages for v3.

Model packages are published separately on Hugging Face and are not duplicated in this GitHub Release.

first Released [v0.1.4]

Unsigned and assembleDebug build.