Whisper Large V3 Turbo vs V3: 5× Faster on Mac (Benchmark)

V3 Turbo vs V3: What Changed

Turbo is not a new architecture. It's the exact same Whisper Large-v3 model with the decoder pruned from 32 layers to 4, then fine-tuned to recover accuracy. The encoder is untouched.

Method: the same 10-minute audio file was transcribed in the same Whisper Notes build on each named device. Times are wall-clock seconds from starting transcription to final text; only the model changed between V3 and Turbo.

The translation task was explicitly excluded from Turbo's training data. The full Large-v3 model supports it, but Whisper Notes ships Turbo rather than full Large-v3 — translation is handled separately via Apple Intelligence on Mac.

The Base Model: What Is Whisper Large-v3?

Whisper Large-v3 is OpenAI's flagship open-source speech recognition model, released in November 2023. It has 1.55B parameters, uses a 128 mel-bin spectrogram input, was trained on 5 million hours of audio (1M weakly labeled + 4M pseudo-labeled), and supports 100+ languages. On the Hugging Face Open ASR Leaderboard it averages ~7.4% word error rate — the accuracy ceiling that Turbo is measured against throughout this article. For how Large-v3 stacks up against every other on-device model, see our Whisper models comparison.

Speed Benchmark: Whisper Notes on Apple Silicon

In Whisper Notes for Mac, Turbo runs via CoreML on the Neural Engine. Processing 10 minutes of audio:

The 5× speedup is specific to Whisper Notes on Apple Silicon, where the smaller decoder benefits from Neural Engine optimization. On GPU with frameworks like faster-whisper, the gap narrows to ~2.7× (see community benchmarks below).

One clarification about the iPhone and iPad rows: those are model-evaluation runs on that hardware, not the model list the shipping iPhone app offers. Whisper Notes for iPhone does not offer Large V3 Turbo — a deliberate choice, explained in the per-device table below.

Accuracy: WER Comparison

The Hugging Face Open ASR Leaderboard tests both models on the same English datasets. Turbo's word error rate is within half a point of V3 across every benchmark:

V3 is slightly more accurate on every dataset, but the gap is small — 0.39 percentage points on average. For most real-world transcription, you won't hear the difference.

On the YouTube-commons long-form evaluation (one of the largest open-source ASR benchmarks), Turbo scores 13.40% WER vs V3's 13.20% — while running at 129.5× real-time factor vs 55.3×. That's 2.3× faster with nearly identical accuracy on real-world audio.

How Accurate Is Turbo in Korean, Russian, and Other Languages?

The benchmarks above are English. Per OpenAI's model card, Turbo's pruned 4-layer decoder costs slightly more accuracy in non-English languages than in English, with the largest degradation on lower-resource languages. For Russian and most European languages, Turbo stays close to full Large-v3 — and if you're on Whisper Notes, Parakeet V3 covers Russian and 24 other European languages at 10× Whisper's speed.

For Korean, Japanese, Chinese, and Cantonese, a purpose-built model is both faster and better-punctuated: SenseVoice transcribes CJK at 52× real-time. Whisper Notes ships SenseVoice on iPhone and in both Mac channels, alongside Turbo on Mac, so you can pick the right model per language instead of forcing everything through one.

Which Local Speech-to-Text Model Should You Use?

Large-v3 Turbo is the best general-purpose Whisper model you can run locally. It is not the best model for most individual jobs, because on Apple Silicon it competes with two specialists that beat it inside their own language ranges. In Whisper Notes the practical lineup is four models — Parakeet V3, SenseVoice Small, Whisper Small, and Whisper Large V3 Turbo — and the choice is decided almost entirely by the language you speak and the hardware you own, not by the leaderboard you read.

Side by side, with every number in this table drawn from our own Apple Silicon tests or the published benchmarks cited below it:

English WER is the mean across the Hugging Face Open ASR Leaderboard's English datasets. SenseVoice Small and Whisper Small were not part of that English run, so we leave those cells empty rather than guess. On the 25-language European FLEURS average the numbers are 12.0% for Parakeet V3, 12.6% for full Whisper Large V3, and 29.8% for Whisper Small; Turbo was not benchmarked separately on that multilingual set. Speed figures are wall-clock from our own Apple Silicon runs, except the two relative multipliers, which are against full Large-v3. Download and memory figures are measured in the Whisper Notes Mac build on an M4 Pro.

The Same Lineup, Right-Sized per Device

Which of these you can actually load depends on which build you are running. Whisper Notes ships on three channels — the iPhone app, the Mac App Store build, and the Mac Direct Download (DMG) — and the model lists are close but not identical:

That last row is a product decision, not an omission. Turbo is a ~1.6 GB download that wants roughly 1.6 GB of memory while it runs — a reasonable ask on an Apple Silicon Mac, an unreasonable one on most iPhones, where the same storage and memory buy you a model lineup that is faster and, in its own languages, more accurate. So the iPhone app ships Parakeet V3 built in and offers Whisper Small and SenseVoice Small as optional downloads: Parakeet for the 25 European languages at 6.32% English WER, SenseVoice for CJK and Cantonese, and Whisper Small as the 101-language fallback that fits comfortably in a phone's memory budget. Right-sized models per device, rather than the same 1.6 GB file on everything that can technically hold it.

Every model above also has a failure mode worth knowing before you commit to it — the honest version is in What Each Model Gets Wrong further down. For the full side-by-side of every on-device option, including models Whisper Notes does not ship, see our Whisper models comparison.

Community Benchmarks: GPU & CPU

Independent benchmarks from the faster-whisper and whisper.cpp communities show consistent results across hardware. Transcribing 13 minutes of audio with faster-whisper on GPU:

Source: faster-whisper benchmark on NVIDIA GPU, LibriSpeech clean validation split. Turbo int8 uses only 1.5 GB VRAM — it fits on a 2 GB GPU.

Batched inference on an RTX 3060 Laptop (6 GB VRAM, int8 precision) pushes the advantage further:

Source: NilaierMusic benchmark, Intel i7-12650H + RTX 3060 Laptop 6 GB, French audio, int8 precision.

With batched processing, Turbo achieves the best WER of any model tested (7.7%) while being the fastest. It's the clear sweet spot for production use.

Turbo vs Medium vs Every Whisper Model Size

Before Turbo, Medium was the usual compromise: acceptable accuracy at tolerable speed. Turbo makes that trade-off obsolete — at 809M parameters it is barely larger than Medium (769M), yet delivers large-class accuracy at several times the speed. Here's the full model family side by side:

Released September 30, 2024, Turbo is 809M parameters. If you were choosing Medium to save disk space or speed, Turbo now beats it on both accuracy and speed at roughly the same footprint.

Whisper Notes ships only two rows of that table — Whisper Small and Whisper Large V3 Turbo. Tiny, base, and medium are omitted on purpose: on Apple Silicon each of them is dominated by something else in the lineup. Anything you would have run tiny or base for, Parakeet V3 does faster and far more accurately; anything you would have run medium for, Turbo does better at a similar size. Whisper Small survives because nothing else covers 101 languages in under 500 MB of memory.

Known Limitations (and How Whisper Notes Handles Them)

What Each Model Gets Wrong

Every model in this lineup has something it is genuinely bad at, and the honest list is more useful than another benchmark table. Whisper — both Small and Large V3 Turbo — hallucinates on silence. Its decoder is autoregressive: it predicts the next token whether or not there is anything left to hear, so a quiet stretch can come back as a repeated phrase, a stray subtitle credit, or a fluent sentence nobody said. Turbo's 4-layer decoder makes this somewhat more likely than full Large-v3 on very short or noisy clips. Voice activity detection removes most of it by never sending the silence to the model in the first place, and Parakeet V3 emitted far fewer invented words in our silence tests — but no speech model deserves a "never" claim, and you should still read a transcript before you rely on it.

Parakeet V3 has a hard 25-language ceiling, and all 25 are European. No Chinese, Japanese, Korean, Cantonese, Arabic, Hindi, Thai, or Hebrew — and it will not warn you when the audio drifts outside its set, it will simply produce confident nonsense. SenseVoice Small is narrower still: English, Simplified and Traditional Chinese, Cantonese, Japanese, and Korean, and nothing else. It is the fastest model here on the material it was built for and the wrong tool for everything else. Whisper Small is the cheapest to run and the least accurate — 29.8% average WER across the 25 European FLEURS languages against Parakeet V3's 12.0%, which shows up first on proper nouns, technical vocabulary, and accented speech.

And all of them cost storage and hardware. These are real files on your disk, not an API call: Turbo is about a 1.6 GB download that uses roughly 1.6 GB of memory while transcribing, SenseVoice Small 827 MB, Whisper Small 600 MB, Parakeet V3 465 MB. The whole lineup runs on Apple Silicon only — macOS 14 or later on M1 and newer, iOS 18 or later on iPhone. There is no Intel Mac build, no Windows build, and no Android build, and nothing here transcribes live: Whisper Notes processes a recording after you stop it, not while you speak.

Whisper Large-v3 Turbo FAQ

Comparing all the local options? Every on-device speech-to-text model — Whisper variants, Parakeet V3, SenseVoice, and Voxtral — is compared side by side on our Whisper models comparison page. New to Whisper itself? Start with the Whisper Transcription Guide — what the model is, every way to run it, and what it costs.

Download for iOS Download for macOS