Comparing Speech-to-Text across Cloud, Edge and Local Deployments

Real-time speech-to-text is the front door of a spoken command-and-control interface. In drone operation, autonomous vehicles and industrial robotics, it has to be both fast and right — a late transcription and a wrong transcription fail in different but equally unacceptable ways, and in a high-stakes environment either one can become an operational failure or a safety risk.

The three ways to deploy it fail differently. Cloud services from Azure and Google bring enormous training data and large models to bear, and pay for it with a network round trip that can run to hundreds of milliseconds — plus a hard dependency on the link being there at all. Offline engines like Silero and Whisper keep the audio on the device, which removes both the latency and the privacy exposure, but bounds you by whatever compute the vehicle can carry. Between them sits what the paper calls the local cloud: a dedicated on-premises GPU server that runs the large accurate model without the internet round trip.

I evaluated all three across a spread of hardware rather than one reference machine, because the deployment question is really a hardware question: Nvidia Orin Nano and Turing RK1 on the edge, and desktop systems with 1080Ti and 2080Ti GPUs for the local case, each tested on CPU and on CUDA where applicable. The pipeline standardizes audio to 16 kHz, runs each sample from the Fluent Speech Corpus five times to capture variability, and reports two metrics: end-to-end elapsed time from audio in to transcription out, and word error rate against ground truth.

The results split cleanly by category. Azure achieved the lowest mean WER at roughly 0.03–0.04 with a tight distribution. Google came in slightly higher at 0.05–0.07 and noticeably broader. Among offline engines, Whisper beat Silero locally — about 0.06–0.08 against 0.10–0.11 — though both showed wider variability than the cloud services. Edge deployments of both engines had the highest error rates and the most variable processing times, which is the cost of running on a board a vehicle can carry.

Latency inverted that ranking. Local GPU inference through CUDA returned transcriptions in 0.2–0.4 seconds, comfortably faster than any cloud path once network overhead was counted, for only a modest increase in error rate.

The conclusion is that neither pole is correct on its own. The trade favors a hybrid: local computation for the immediate response that the control loop needs, with cloud services available for accuracy where the extra latency is affordable. Which side a given command falls on is an engineering decision, and this is the data to make it with.