Starstruk Logo MarkStarstruk

v1.9.0

FranckyB/Voice-Clone-Studiov1.9.0Mar 1, 2026by FranckyB

AI Summary

This release brings major performance improvements through Faster-Qwen3-TTS integration with CUDA graph acceleration (5-10x faster inference) and introduces multi-GPU support for distributing TTS, ASR, and Llama.cpp workloads across different GPUs.

Key Highlights

  • Faster-Qwen3-TTS integration with 5-10x faster CUDA graph-accelerated inference
  • Multi-GPU support with GPU assignment dropdowns for TTS, ASR, and Llama.cpp
  • Per-subsystem GPU control to maximize throughput
  • Toggle in Settings to enable/disable CUDA Graphs Acceleration
  • Automatic fallback to standard Qwen3TTSModel when CUDA graphs unavailable
  • Trained model caching between generations
  • Setup scripts auto-install Faster-Qwen3-TTS package

New Features

  • CUDA Graphs acceleration for Qwen3 generation
  • Multi-GPU workload distribution
  • Configurable GPU assignments saved to config.json
  • Trained model checkpoint caching

Full Release Notes

**Faster-Qwen3-TTS Integration**
- **5-10x Faster Inference** - Integrated [Faster-Qwen3-TTS](https://github.com/andimarafioti/faster-qwen3-tts) for CUDA graph-accelerated Qwen3 generation with bit-identical output quality
- **All Qwen3 Models** - Acceleration applies to Base, CustomVoice, VoiceDesign, and Trained Model checkpoints
- **Toggle in Settings** - Enable/disable CUDA Graphs Acceleration under Faster-Qwen3-TTS section (CUDA only)
- **Automatic Fallback** - Gracefully falls back to standard Qwen3TTSModel when CUDA graphs are unavailable or fail
- **Setup Scripts Updated** - `setup-windows.bat` and `setup-linux.sh` auto-install the package (not on macOS)
- **Trained Model Caching** - Trained model checkpoints are now cached between generations instead of reloading every time

**Multi-GPU Support**
- **GPU Assignment Dropdowns** - Assign TTS, ASR, and Llama.cpp to different GPUs on multi-GPU systems
- **Per-Subsystem Control** - Each subsystem (TTS, ASR, LLM) can run on a separate GPU to maximize throughput
- **Automatic Detection** - GPU dropdowns only appear when multiple CUDA GPUs are detected
- **Saved Preferences** - GPU assignments persist across restarts via config.json

**Bug Fixes**
- **Conversation Tab Fix** - Fixed a broken component reference (`vv_conv_sentences_per_chunk` → `vv_conv_paragraph_per_chunk`) that prevented all interactive elements from working on the Conversation tab