Whisper vs AssemblyAI — Which Transcribes Portuguese Better?

On clean audio with a single speaker: accuracy is practically identical (94-97%) for PT-BR. Differences appear in specific cases: 1) Noisy audio: Whisper large-v3 has an edge (trained specifically for noise robustness). 2) Multiple speakers: AssemblyAI has an edge (native diarization). 3) Automatic punctuation: AssemblyAI inserts periods, commas, question marks. Whisper produces flowing text without punctuation. 4) Regional Brazilian accents: similar, with a slight AssemblyAI advantage in internal tests.

Open-source Whisper is free but not simple to use: requires Python installed, GPU for acceptable speed (CPU is 10-20x slower), and technical configuration. Via OpenAI API, Whisper costs $0.006/min — cheap, but without diarization, automatic punctuation, or advanced features. AssemblyAI (and VoxScriber which uses it) delivers everything ready: web interface, export formats, diarization, punctuation — no technical setup.

AssemblyAI is async — send the file and receive results when ready (no chunking needed for files up to 5GB). For a 1-hour file: AssemblyAI returns in ~3-4 minutes. Whisper via OpenAI API is synchronous but with a 25MB limit — larger files need manual chunking (split + merge), adding complexity and time. For production with large files, AssemblyAI is significantly simpler.

VoxScriber uses AssemblyAI as the default engine for all users (free and paid). It is the engine with the best accuracy-to-features ratio for Portuguese: native diarization, automatic punctuation, and no chunking needed for large files. It costs 4 cycles/minute. Users on LITE and Advanced plans can switch to Whisper (cheaper, 1 cycle/minute) for specific cases such as very noisy audio.

For interviews with 2 speakers and medium audio quality (phone recorder or Zoom): AssemblyAI Best + diarization is the best choice — automatically identifies speakers, adds punctuation, and has precise per-sentence timestamps. Whisper large-v3 produces more accurate text in bad audio, but without speaker identification — you would need to use pyannote.audio in post-processing for diarization.

Yes. In VoxScriber (LITE+ plans), you can choose the engine per file. Common workflow: use AssemblyAI for standard files (meetings, interviews with good quality) and switch to Whisper when a file has a lot of background noise or low quality. Credits are debited per engine: AssemblyAI = 4 cycles/min; Whisper = 1 cycle/min (cheaper).