Qwen3-ASR and Kokoro-TTS ONNX on CPU enable GPU-free voice assistants

A user tested the latency of Qwen3-ASR and Kokoro-TTS running as ONNX models on CPU to determine how much processing load can be offloaded from the GPU in a voice assistant pipeline.

• The test utilized Daumee/Qwen3-ASR-0.6B-ONNX-CPU and onnx-community/Kokoro-82M-v1.0-ONNX models.
• Latency was measured on a 2022 MacBook M2 and an AMD Ryzen 9 7900, with the latter described as "blazing fast" while the M2 remained mostly usable.
• The setup employed a 5-second follow-up window and Voice Activity Detection (VAD) to trigger regex commands without requiring a wakeword for every interaction.

This approach frees up the GPU entirely for running the Large Language Model, potentially opening up new possibilities for local voice assistant architectures.