The audio has been resampled to mono 16‑bit PCM @ 8 kHz to match the ASR input.
Recognition is performed phrase by phrase. Make a short pause for the text to appear in the list.
Experimental implementation. The demo runs entirely on a CPU to keep the setup portable. For production‑grade latency and throughput we recommend running the models on a GPU and, where possible, applying runtime optimisations such as TensorRT (not included in the demo).