The microbenchmark answered a narrow question well
Eighteen text-to-speech models were tested on one Apple Silicon machine. The fastest warm run averaged about 0.19 seconds for a short fixed input after the model was already loaded. That number describes steady-state synthesis under one workload. It does not describe what a person hears when they ask the product to speak.
Hardware, quantization, text length, voice, sampling settings, and warmup state all constrain the result. Those conditions should be published beside the number. Otherwise a local measurement escapes into a general product claim, usually in a slide that says the system can do real-time voice.
The live path contained the other eight seconds
The production pipeline added model selection, process startup, weight loading, text normalization, sentence segmentation, audio encoding, buffering, and transport. Some stages were sequential. Some repeated work the microbenchmark had already warmed. First audible output arrived around nine seconds after the request.
| Measured boundary | Elapsed seconds |
|---|---|
| Warm synthesis | 0.19 |
| Request to first audible output | 9 |
Both numbers were true. The model was fast once it was ready. The product was slow at the boundary that mattered to a listener. Further optimization of synthesis would have been a satisfying way to miss most of the delay.
Instrument the latency ladder, then assign owners
Record request receipt, model-ready time, first synthesis chunk, first encoded chunk, first byte sent, and first audio played. Report cold and warm distributions separately, including p50 and p95, rather than one best run from a warm cache.
The ladder assigns each delay to an owner. Model caching affects readiness. Streaming synthesis affects the first chunk. Encoding and buffering affect delivery. Client playback affects perceived start. Without those timestamps, the argument collapses into a claim that the model is slow, and the team buys a different model for a pipeline problem.
Optimize for first useful output
For interactive speech, time to first audio often matters more than total synthesis time. Preloading a frequently used model, segmenting text early, streaming chunks, and keeping the audio path warm can improve the experience without changing the model.
A slower model may still win if it streams earlier or produces audio that needs less post-processing. The selection metric should match the product boundary: quality, first-audio latency, total latency, memory footprint, and failure recovery on the target hardware. The 0.19-second figure is a component measurement. The nine-second figure is the one a listener experiences.