A specific, documented complaint about the category leader: speak in one language, and the app spontaneously translates to another without being asked. It's a real bug, named by real users, and worth understanding structurally before picking a dictation app for a multilingual workflow.
Two separate, dated organic complaints describe the same failure mode in Wispr Flow: "I speak English (mostly) and Portuguese… Wispr started to translate from one language to the other… After the ~20th time, I gave up and disabled Portuguese" (2026-06-15), and "Wispr Flow live translates when I speak Swedish into English. As far as I know, live translation is not even a feature" (2026-06-30). Wispr Flow's auto-detect was also removed at some point, which reportedly angered European users who relied on it to switch languages mid-session without manual configuration.
This isn't a claim that every cloud dictation product has this exact bug — it's a specific, sourced complaint about one product. It's included here because it points at a structural question worth asking about any dictation app you're evaluating for multilingual use: what happens between what you said and what lands on the page, and how many opportunities are there for that pipeline to do something you didn't ask for?
OpenWispr's on-device pipeline runs a speech model — NVIDIA Parakeet-TDT by default, or Whisper via whisper.cpp — directly on your device, with no network round-trip that could route audio through a different model, service version, or configuration than the one you started with. That's a structural property of the architecture, not a benchmarked claim about translation accuracy specifically — this page isn't asserting OpenWispr has been tested head-to-head against the exact failure mode described above, only that there's no cloud-side model swap or hidden mode change in the pipeline to produce it.
For context on language breadth elsewhere in this category: Muesli publishes a model table showing Parakeet v3 covering 25 languages and Nemotron 3.5 covering 100+ locales, on-device — a useful reference point for how far local models have come on multilingual coverage generally, separate from the specific translation bug discussed above.
Open source and entirely on your device. Android from Google Play, macOS from GitHub — no account, no sign-up, just a download.