Whisper large-v3-turbo vs large-v3: the smaller model won — then the full one looped 14 times

By Blasting Reviews (independent tester, EU) · Published Nov 8, 2026 · Updated Nov 10, 2026 · Corrections & tips: hello@blasting-reviews.com

Bench resultUse itverdict4.39 %FR word error rateApple-silicon bench · method · all results

whisper.cpp turbo: 4.39% FR / 1.83% EN WER — MLX Whisper: 10.12% FR on identical audio.

turbo-vs-full-whisper-bench

Review · speech-to-text bench, part 2 — measured Nov 8, 2026

Use it

Run transcription locally with whisper.cpp + large-v3-turbo: in our bench the 8×-smaller turbo model beat the full large-v3 on accuracy and was 3–4× faster, for €0 and zero internet.

The conventional wisdom says the small model is the compromise: you trade accuracy for speed. We tested that trade on our transcription bench — same audio, same scoring method as our Nemotron-vs-Whisper review, two ggml models in whisper.cpp — and the trade didn’t show up. The full model lost both races.

Model A: whisper.cpp + ggml-large-v3-turbo (0.8 B params, 1.6 GB pack) Model B: whisper.cpp + ggml-large-v3 (1.6 B params, 3.1 GB pack) Hardware: Apple M5 Ultra, Metal backend, same machine back-to-back Audio: ~10-min Alice in Wonderland chapter (EN) · ~6-min Jules Verne chapter (FR) Scoring: word-by-word Levenshtein against the printed book

The numbers

ModelEnglish WERFrench WERTime EN clipTime FR clip
large-v3-turbo1.83 %4.39 %11.5 s7.5 s
large-v3 (full)10.92 %6.36 %40.5 s24.4 s

Same machine, same runtime, same decoder settings. The smaller, 3.5× faster model transcribed both languages with fewer errors. Case closed? Not quite — and the asterisk matters.

What the 10.92 % actually is: a captured repetition loop

Word Error Rate punishes hallucinated words twice: they count as insertions, and they bury the real text. That is exactly what happened. Eleven minutes into Alice, the full model stopped following the reading and started spinning:

“…because she thought it was a bad thing to do, and she thought it was a good thing to do, and she thought it was a bad thing to do…” — that clause cycle ran 14 times, verbatim.

Our diff caught 239 edit operations against the printed text, most of them generated by that loop; by the end of the chapter the model had recovered and finished cleanly (“…very soon finished off the cake”). So the honest reading of this bench is not “large-v3 is 6× worse than turbo”. It is:

  • large-v3 has a real, documented-in-practice repetition failure mode, and on a 10-minute clip the failure cost us a 6× WER penalty with no warning in the output.
  • large-v3-turbo did not hit a single loop in either language — 40 and 49 total edits, all plausibly real misrecognitions.
  • Even excluding the loop, the full model was worse in French too (6.36 % vs 4.39 %) with no failure mode to blame. On audiobook-style speech, this is a genuine loss.

When large-v3 still earns its 3 GB

We are not claiming large-v3 is obsolete. It carries non-turbo features the turbo pack dropped — translation from other languages, and the full multilingual training that helps on noisy, accented or specialist audio. Our clips were clean studio readings. If your corpus is messy meeting recordings, the round-2 noisy-audio bench we have planned is precisely the test that could reverse this ranking; we will update this table, not quietly edit it.

Practical takeaways

  1. Default to turbo. 11.5 seconds for a 10-minute chapter (~52× real time) with the best accuracy we measured locally. That is not a compromise, it is the current sweet spot.
  2. Validate length, not just text. The loop added ~50 words of junk; a silent omission in another runtime dropped ~70 (our next post). Two cheap sanity checks catch both: word count vs expected, and repeated-phrase detection. Two lines of script, worth it in any pipeline.
  3. Benchmark on your own audio. Every number in the industry press is somebody else’s corpus. Two chapters and a Saturday afternoon is enough to know which model wins on yours.

Method, exactly

Whisper.cpp (Homebrew, Metal backend) with the official ggml quantization packs, default 5-second-beam decoding, fp16 on Apple silicon. Ground truth from Project Gutenberg and Wikisource; punctuation and case stripped, Unicode normalised, then global Levenshtein alignment; WER = (substitutions + insertions + deletions) / reference words. Raw transcripts, reference texts, and the scoring script are in our repo; the /benchmarks/ page links the versioned results.

Use it

For local transcription today: whisper.cpp + large-v3-turbo. It won on speed, on English accuracy, and on French accuracy — and it never lost the plot. The full model keeps its place only where its extra features are needed.

Get the next benchmark in your inbox

One email per measured test — the numbers, the method, the raw data. Twice a month at most, no third-party trackers.

By subscribing you agree to receive occasional emails from Blasting Reviews. One-click unsubscribe in every email; the list lives on our own EU server. Correct anything: hello@blasting-reviews.com.

Cite as: Blasting Reviews (2026-11-08). “Whisper large-v3-turbo vs large-v3: the smaller model won — then the full one looped 14 times” · 4.39 % FR word error rate · Apple M5 Ultra bench · BR-STT-2026-11-08-75 · https://blasting-reviews.com/whisper-large-v3-turbo-vs-large-v3-audiobook-bench

Blasting Reviews

Independent tests of software, smart home gear, gadgets and everyday tech — run on our own bench, with the method and the numbers published. Verdicts stay ours; purchase links never change them.

Sections

AI & Models
Apps & Productivity
Streaming & Media
Smart Home & Energy
Gadgets & Hardware
Privacy & Services

Company

Benchmarks
About
How we test
Contact

Fair play

Some links may be affiliate links — they never influence a verdict, and prices are dated on publication.