Whisper models: which size to run, and Large V3 or Turbo

OpenAI released Whisper as open weights under the MIT license, trained on a very large amount of weakly labeled audio from the web. That training is why it copes well with accents, background noise and many languages, and why nearly every local transcription app on the Mac offers it. The hard part is choosing among the Whisper models, because the names say little about the trade-offs.

Whisper model sizes

Parameters, memory and relative speed from the OpenAI Whisper README. Speed was measured by OpenAI on English speech on an A100 GPU, with Large as 1x. Language count from the Hugging Face model cards. These are not Mac timings.

Measured on a Mac

In our own test on an M4 Pro, Whisper Tiny transcribed 8 minutes of clean English in about 6 seconds with a 10.2% error rate. Large V3 took about 100 seconds and cut errors to 3.4%. That gap is why Tiny is fine for a quick draft and Large V3 (or Turbo) is the one to use when the transcript matters.

Our own test, run October 17, 2026 on a Mac with an Apple M4 Pro chip and 24 GB of memory, macOS 27.0.1. Audio: 73 clean read-speech clips from the LibriSpeech validation set (hf-internal-testing/librispeech_asr_dummy on Hugging Face), joined into one 8-minute file, 1,150 reference words. Each model was run twice through the TranscribeX command line and the second run is shown; times are end to end, including loading the model. Word error rate uses simple normalization (lowercase, punctuation removed). Clean read speech is easier than a real meeting, so expect higher error rates on your own calls, and treat differences under one point as noise on a sample this size. Qwen3 ASR was not timed in this round; Whisper Large V3 Turbo, Parakeet V3 and Nemotron were not part of it.

The memory figures are for OpenAI’s reference implementation on a GPU. Mac apps run optimized versions, so the numbers on your machine will differ. They are still a fair guide to which sizes are heavy. In TranscribeX’s model list, Large V3 is a 3.1 GB download and Tiny is 76 MB.

Whisper Large V3 vs Large V3 Turbo

Whisper Large V3 came out in November 2023. Compared with Large V2 it uses 128 mel frequency bins instead of 80, adds a Cantonese language token, and was trained on 1 million hours of weakly labeled audio plus 4 million hours of pseudo-labeled audio. OpenAI reported 10% to 20% fewer errors than Large V2.

Whisper Large V3 Turbo followed in October 2024. OpenAI cut the decoder from 32 layers to 4 and fine-tuned the result for two more epochs. It was not distilled from Large V3. The decoder is the part that writes out text token by token, so shrinking it speeds things up a lot while the encoder, which listens to the audio, stays the same size.

WER and RTFx from the Open ASR Leaderboard paper (Srivastav et al., arXiv 2510.06961, Table 3). RTFx is measured on datacenter GPUs and only compares models with each other. AMI figures from the Hugging Face model card evaluation data. Language notes from OpenAI’s Turbo announcement.

For English, the gap between the two is under half a point of WER, which you will rarely notice. For meetings in English, Turbo is the better daily choice. Pick Large V3 for languages where Turbo is known to slip, for difficult audio, or when you want Whisper to translate speech into English.

Which Whisper model to pick on a Mac

• Apple Silicon Mac, mixed languages: Large V3 Turbo for everyday use, Large V3 when accuracy matters more than time.
• Older Intel Mac or limited memory: Small or Base. TranscribeX works on most modern Macs, but Apple Silicon is much faster because the models use the Neural Engine.
• Notes to yourself, clear audio, no budget: Tiny, free and unlimited.
• English only and lots of audio: consider NVIDIA Parakeet instead. It scores better than Whisper on the English leaderboard and runs far faster.
• Chinese, Japanese or Korean: compare Whisper with Qwen3 ASR on a sample of your audio.

Is Large V3 the best Whisper model? On accuracy, yes. In daily use, the best Whisper model is usually the largest one that finishes fast enough that you don’t mind waiting. With TranscribeX you can switch per file, so try Turbo first and move up if the transcript needs it.

Fine-tuned Whisper models for your language

Whisper’s accuracy varies a lot between languages. Researchers and communities have fine-tuned Large V3 on single languages, which often helps with spelling, names and dialect. TranscribeX Pro includes fine-tuned Whisper models for German, Swiss German, Chinese, Korean, Arabic, Slovak, Catalan, French, Hebrew and Hinglish, with more added on request. If your language is missing, you can ask for it.

A fine-tuned model is only as good as its training data, so treat it like any other model: run a recording you know well through both the general and the fine-tuned version and compare.

Whisper in TranscribeX

TranscribeX runs Whisper from Tiny through Large V3 and Large V3 Turbo locally. Tiny is in the free version; the other sizes come with Pro. Every transcript can then get speaker labels, an AI summary from a local model, translation and exports to SRT, VTT, PDF and more, with the audio staying on your Mac. If you already use MacWhisper, the same Whisper models give essentially the same text; the MacWhisper comparison covers the rest. For a free start, see free transcription software for Mac.