A Hacker News user is seeking the best local, open-source speech-to-speech setup for a true voice assistant, requiring low latency and interruptibility. They found Qwen3 Omni promising but lacking reproducible guides for real-time local deployment, highlighting a gap between research and practical tooling.

The community's response reveals the current state is dominated by piecing together streaming ASR + LLM + TTS pipelines rather than true end-to-end models. Developers cite tools like Pipecat for orchestration and models like NVIDIA Parakeet, Mistral, and Pocket TTS that fit on consumer GPUs, with MLX pipelines noted for Mac users.

Experts confirm end-to-end speech models are still nascent, often underperforming against composite stacks for complex tasks. The hunt continues for a turnkey, low-latency solution on a single GPU, with Kyutai's work and frameworks like LiveKit frequently mentioned as promising foundations for future development.