ExecuTorch expands cross-platform voice models

This is the kind of infrastructure update that matters because voice products die on deployment friction, not model demos, and ExecuTorch is attacking that problem directly.

• –Official announcement: [PyTorch blog post](https://pytorch.org/blog/building-voice-agents-with-executorch-a-cross-platform-foundation-for-on-device-audio/) details Parakeet TDT, Voxtral Realtime, Whisper, Sortformer, and Silero VAD.
• –ExecuTorch homepage: [ExecuTorch](https://executorch.ai/) frames the stack as cross-platform on-device AI with broad backend coverage.
• –The cross-backend story is the real value: XNNPACK, Metal, CUDA, Vulkan, and Qualcomm mean one exported model can reach desktop and mobile without a rewrite.
• –The C++ layer matters as much as the model export, because voice apps need streaming windows, timestamp extraction, caching, and stateful decoding.
• –LM Studio shipping transcription on top of ExecuTorch is the best proof point, but the next credibility test is filling obvious gaps like TTS, live translation, wake-word detection, and noise suppression.