What the no-upload trade-off actually costs
A speech model that fits in a browser is smaller than one running on a server rack. Here is the size of that gap, measured on the same audio, for every language.
word error rate, lower is better · FLEURS test · n = 111–120 · 2026-09-26
How this was measured
Both systems were given the same 150 sentences per language from Google’s FLEURS test set, on the same machine, and scored with the same normalisation: case and punctuation ignored. A sentence is excluded entirely if its reference text, or either system’s output, contains digits — “1940” and “nineteen forty” are both correct, and no single rule normalises that across eleven languages. The same sentences are dropped for both systems, and dropping them widens the gap rather than narrowing it. The reference system is Whisper large-v3, the model most free online transcription services run, quantised to q5_0. Many free services actually run a smaller Whisper variant, so the gap shown here is the widest you should expect.
What these numbers do not tell you
FLEURS is read speech recorded in quiet conditions, so both columns are optimistic compared with a real meeting recording. Read the gap between the columns, not the absolute values. FLEURS also ships only Latin-American Spanish and Brazilian Portuguese, so the European variants of those two are not measured separately.
The last column is the lowest word error rate published for each language on FLEURS, by the system’s own developers or in a research paper — we did not measure it. Those figures use the full test set and stricter scoring than ours: the same Whisper large-v3 scores 2.9 % in German here but 4.5–4.9 % as published. On equal terms, the gap to the best cloud systems is a little wider than these columns suggest.
ElevenLabs Scribe v1 (self-reported) · Mistral Voxtral Mini Transcribe V2 (Voxtral Realtime paper, Table 7) · Whisper large-v3 (OpenAI, large-v3 release) · NVIDIA Canary-1B-v2 (Canary-1B-v2 paper, Table 11)
Measured on this page in Chrome on Apple M2 (8 cores), with 3 recognizers working in parallel as the site does by default: about 20 times faster than real time. Computers with fewer cores, phones and tabs left in the background are slower.