On standard American English, the difference is real but not dramatic. OpenAI's published numbers on the Whisper large-v3 Hugging Face model card and the original paper "Robust Speech Recognition via Large-Scale Weak Supervision" show that large-v3 sits well below medium on English WER across both FLEURS and Common Voice evaluations. The model card lists large-v3 around the mid-2% range on FLEURS English, while medium lands noticeably higher.

That gap is meaningful for production dictation. At 2-3% WER you mostly see punctuation and the occasional homophone ("their" vs "there"). At 6-8% WER you start losing content words, especially names and technical terms. Medium is not broken — it's just noticeably noisier.

The catch: "standard American English in a quiet room" is a narrow slice of what people actually dictate. Move the speaker, the mic, or the accent, and the gap widens.

Honest answer: nothing per-accent. OpenAI evaluates Whisper on whole-language benchmarks — FLEURS (102 languages), VoxPopuli (14 European languages), Common Voice, and LibriSpeech. None of these are broken down by "Indian English," "Brazilian Portuguese-accented English," or "Cantonese-accented English." That's a real gap in the public literature, and I'm not going to paper over it with numbers I made up.

What we can infer from the published data is qualitative. Accents typically degrade WER because they introduce phonetic patterns the model saw less of during training. Models with more parameters and more training data — i.e., large-v3 — generally handle those patterns better. Independent researchers have confirmed this trend in third-party studies, but those are not OpenAI's numbers.

So the honest picture is: large-v3 probably beats medium more on accented English than on standard English, by a margin we can't quantify from OpenAI's published charts alone. That's the truth. If a vendor claims a specific percentage improvement on "Indian accent," ask for the test set.

Yes — and this matters more than the accent question for many users. OpenAI's FLEURS table shows that the WER gap between large-v3 and medium grows substantially as you move from high-resource languages (English, Spanish, French) toward lower-resource ones (Swahili, Welsh, Lao).

On a low-resource language, medium can post double-digit WER while large-v3 stays in single digits. If you dictate in anything other than English — even occasionally — that gap is the strongest argument for large-v3.

This is also why I default to the largest model that fits my hardware in my offline dictation setup. MetaWhisp itself ships with large-v3-turbo specifically because it preserves most of large-v3's accuracy on non-English languages while running in roughly half the time on the Neural Engine.

Noise is where the model-size gap shows up most clearly in practice. OpenAI's Common Voice and VoxPopuli evaluations both include varied recording conditions, and large-v3 dominates medium on noisy clips. Large-v3 also hallucinates noticeably less on silence and background hum — the famous Whisper problem where the model invents a sentence during a quiet stretch.

Medium hallucinates more. If you've ever finished dictating and found a phantom "Thanks for watching!" at the end of your transcript, that's medium's classic failure mode. Large-v3 doesn't eliminate it, but it's rarer and usually shorter when it does happen.

For long-form dictation — meetings, lectures, podcasts — hallucination matters more than a 2% WER improvement. Every hallucinated sentence is cleanup work. This alone pushes the value proposition toward large-v3 for anything beyond quick voice notes.

Yes, with caveats. Whisper large-v3 runs on Apple Silicon via WhisperKit, and on an M1 or later it will produce a transcript — but the latency depends heavily on the chip and audio length. On an M1 Air with the official Core ML port, large-v3 runs slower than medium — a 60-second clip takes a couple of minutes to transcribe.

An M2 Pro or M3 Pro cuts that significantly. On M4 Pro and Max chips, large-v3 feels nearly real-time for short clips. For an M1 Air user who dictates a lot, large-v3-turbo is the practical sweet spot — about half the latency of large-v3 with accuracy that the model card and third-party tests put within roughly 1 percentage point of the full model on most languages.

If your workflow is "hold hotkey, dictate three sentences, paste into email," any current Apple Silicon Mac handles large-v3 fine. If you're transcribing an hour-long meeting, you'll want either a Pro chip, large-v3-turbo, or a cloud pass to skip the wait.

large-v3-turbo is OpenAI's distilled variant of large-v3 — fewer decoder layers, similar encoder, same general accuracy tier. It launched in late 2024 and is what most production apps now ship by default. The tradeoff it makes: a small accuracy drop (typically under 1 percentage point WER on FLEURS) for a large speedup, often 2x or better on the same hardware.

For accented English specifically, large-v3-turbo is the option I usually recommend. It keeps nearly all of large-v3's gains over medium while running in the time budget of an everyday dictation app. MetaWhisp uses it for exactly this reason — local mode is free and unlimited on macOS 14+ Apple Silicon, and the accuracy hit versus the full large-v3 is small enough that most users don't notice.

I wrote a deeper dive on large-v3-turbo vs full large-v3 if you want the model-level details.