Why not local Whisper
The case for server-side transcription in a native app.
Local Whisper is appealing. No network dependency, no API costs, no latency from a round trip. In theory, it is the obvious choice for a voice transcription app.
In practice, the large-v3 model is 1.5 GB. On Android, loading it into memory competes with every other app. On Linux laptops, it pins a CPU core and drains battery. On low-end Windows machines, it simply does not run fast enough to feel responsive.
Groq Whisper returns results in under two seconds on a standard connection. The model runs on dedicated inference hardware that stays cool, stays fast, and does not drain anything on the user's device.
The tradeoff is real: you need an internet connection. But for the target use case — speaking into a text field while working — you already have one. The network is not the bottleneck. The user experience is.