On-device speech AI has left the lab. The receipt is the engineering tax.

On-device speech AI has stopped being a research demo. The phones are doing the work now, and the more interesting fact is what that crossing cost. Li's open-source iPhone app, LiveTranscriber, lists five working engines: Whisper, Qwen3-ASR, NVIDIA Nemotron Streaming, MOSS Multi-Speaker, and Qwen3 for summaries. Li's own README names the bill: memory pressure, streaming latency, model loading, context handling, battery drain, and switching between inference backends. The popular read is that phones can now do speech AI. The engineering-tax framing is a reporter interpretation. Individual model performance, actual model-swap latency, and real-world battery impact have not been independently benchmarked. The pattern underneath is sharper. Li's privacy policy reveals the five open-source models are selectable engines, not a stacked live pipeline, so the real cost a user pays is model-swap latency and multi-backend context, not five models running in parallel. Works offline, in this cycle, is an engineering choice, not a default. The next on-device speech story that lands on the desk will repeat the same shape: a maker ships a working pipeline, names the tax, and quietly proves that "free of the cloud" is a labor bill someone has to pay.