Spanish transcription benchmark
Why does AudioMap use AssemblyAI and not Whisper? Because every percentage point of Word Error Rate (WER) shows in the final result. Here is the public data by engine and regional variant.
WER (Word Error Rate): percentage of incorrect words out of total. Lower = better. 0% would be perfect. For professional use (legal, medical, journalism) WER <10% is considered acceptable; ideal <7%.
The figures in the table are for AssemblyAI Universal-2, the previous generation. In production we run Universal-3.5 Pro, for which AssemblyAI does not publish a comparable WER by regional variant.
AssemblyAI wins in all variants
Universal-2 (previous generation) maintains WER <11% across all 5 tested Hispanic variants. Whisper-large-v3 ranges from 14.7% to 19.1%. A 7-10 point difference in regions with yeísmo or high speech rate (Rioplatense, Chilean).
Whisper is reasonable only in neutral Spanish
Whisper performs acceptably in standard es-ES but loses quality in regional variants. TurboScribe, Fireflies (non-English) and many low-cost alternatives use it, you are paying for quality that is not the best available.
Coming: AudioMap own benchmark
Q3 2026 we will publish a benchmark with our own curated audio (10h per regional variant, professional domain, human evaluation). We promise transparency: open dataset and methodology.
300 minutes per month included. No credit card.