Whisper Large V3 Turbo
Whisper Large V3 Turbo is OpenAI's Whisper with the decoder cut from 32 layers to 4, which makes it several times faster than the full Large model at a small cost in accuracy (7.8% against 7.4% average English WER). FreeFlow runs it as a 1.6 GB CoreML model through WhisperKit. It covers 99 languages, Hebrew included, and can translate what you say into English.
| Full name | OpenAI Whisper large-v3-turbo (openai_whisper-large-v3-v20240930_turbo) |
|---|---|
| Download in FreeFlow | 1.6 GB |
| Languages | 99, with automatic detection |
| Average English WER | 7.8% (Open ASR Leaderboard, checked 2026-11-09) |
| Speed on Apple Silicon | A second or two for a sentence; slower than Parakeet, far faster than Whisper Large |
| Runs through | WhisperKit, Argmax's open-source CoreML runtime |
| Translate to English | Yes |
| Needs | macOS 14+, Apple Silicon |
When to pick it
- Hebrew. Neither Parakeet nor Cohere Transcribe has it; this is the best Whisper for it in FreeFlow.
- Any of Whisper's 99 languages outside Parakeet's 25 and Cohere's 14 — Hindi, Turkish, Indonesian, Thai, Persian and so on.
- Translate to English as you dictate: speak your language, English is typed. A per-mode switch; see languages.
- Mixed-language dictation where the speaker switches mid-sentence.
Turbo against the full Large
The full Whisper Large V3 is 3.1 GB and FreeFlow's slowest model; Turbo keeps its encoder and the same training but trims the decoder, trading 0.4 points of English word error rate for a large speed gain. For dictation — short takes, read immediately — Turbo is the better fit; Large is for someone who wants the most thorough Whisper and does not mind the wait. Whisper Small (970 MB, about 8%) is the light option when disk matters more than accuracy.
Known habits
- Hallucinated phrases on silence — Whisper's famous “Thank you.” for a second of room noise. FreeFlow's silence gate looks for actual speech-level sound before transcribing, so an empty hold pastes nothing.
- Slightly lower English accuracy than Parakeet. 7.8% against 6.3%; on technical English with a good dictionary the gap is smaller than it reads.
- Autoregressive decoding means long takes cost proportionally more time than on Parakeet.
Bring a variant
Fine-tunes and distilled variants of Whisper converted to WhisperKit's CoreML layout can be added from Hugging Face and assigned to any mode.
Questions people ask
Is Whisper Turbo as accurate as Whisper Large?
Nearly: 7.8% against 7.4% average English WER, at a fraction of the time. For dictation the speed is worth more than the 0.4 points.
Which Whisper is best for Hebrew in FreeFlow?
Large V3 Turbo. Full Large is marginally more thorough and twice the download and wait.
Does Whisper send anything to OpenAI?
No. It is a CoreML model on your Mac; the only network use is the one-time download.
Can it translate into languages other than English?
No. Whisper translates into English only.