⚡ faster-whisper vs OpenAI Whisper: 4x Speed, Same Accuracy?

SYSTRAN's own README says faster-whisper runs "up to 4 times faster than openai/whisper for the same accuracy." Its own benchmark table, a few paragraphs below that line, doesn't show a single row hitting 4x. The real, directly comparable gap in that table runs 2.27x to 2.66x. Still an easy call. Just not the headline number.

By the numbers

• GPU, large-v2 model, 13-minute test file, fp16, beam size 5: faster-whisper finishes in 1m03s against openai/whisper's 2m23s, a measured 2.27x, not 4x (SYSTRAN/faster-whisper)
• Same GPU test switched to int8 quantization: faster-whisper drops to 59s using 2,926MB of VRAM, down from 4,708MB, a 2.42x gain against stock with 38% less memory along the way
• CPU, 8 threads, small model, fp32, beam size 5: faster-whisper finishes in 2m37s against openai/whisper's 6m58s, a 2.66x gain, the best directly comparable number anywhere in their own table
• Both tools decode identical published model weights (openai/whisper), so there's no independent word-error-rate gap to measure between them at a given model size

Where does "4x" actually come from, then?

Quoting a benchmark table's best row as the headline is a bit like a runner citing their fastest split as their marathon pace. The one row in SYSTRAN's README that gets close to matching the marketing line isn't even a faster-whisper-vs-openai-whisper comparison. It's batched faster-whisper against the Hugging Face Transformers pipeline on distil-whisper-large-v3, a different model running through a different library entirely (25m50s vs 46m12s, with faster-whisper actually scoring a better WER: 13.527 against 14.801). That's a real result. It's just answering a different question than the one most people read the "4x" line as answering.

A third-party comparison running a full one-hour file across several hardware tiers (CPU-only, an RTX 3090, an A100, an Apple M2) reported a wider band — roughly 4x to 6.7x depending on the tier. That's closer to the marketed number, and worth knowing about. It's also a single outside source rather than the vendor's own controlled numbers, so I'm treating it as directional support, not proof the headline holds everywhere.

Accuracy: same weights, so where do people actually see a difference?

When someone tells me faster-whisper "feels less accurate," it's almost never the CTranslate2 conversion talking. It's a beam-size mismatch they didn't notice they introduced, or a jump to a smaller or distilled model in the same breath as the tool swap. The engine change alone doesn't touch the weights.

The platform catch nobody mentions

faster-whisper has no Apple Metal acceleration. On a Mac it runs CPU-only, full stop. If your machine is Apple Silicon, whisper.cpp is the better-fitting tool, since it taps Metal and Core ML directly. Recommending faster-whisper blind to every platform is exactly the kind of unquantified "just use this" advice I'm trying not to hand you here.

Batching: where the bigger numbers actually live

The distil-whisper row above hints at the real lever behind the wilder speedup claims floating around online: batching. Feed faster-whisper's BatchedInferencePipeline a queue of audio instead of one file at a time, and throughput climbs well past the single-file numbers in the basic benchmark table, which is also closer to where a genuine 4x-or-better result tends to show up in the wild. That's a real workload difference, not a hardware difference. A single podcast episode transcribed alone and a folder of 200 files transcribed together are not the same benchmark, even on identical silicon.

The table, in full

Test (SYSTRAN's own README)	openai/whisper	faster-whisper	Measured speedup
GPU, large-v2, fp16, beam 5, 13-min file	2m23s (4,708MB)	1m03s (4,525MB)	2.27x
GPU, large-v2, int8, beam 5, 13-min file	2m23s (4,708MB)	59s (2,926MB)	2.42x
CPU (8 threads), small model, fp32, beam 5	6m58s (2,335MB)	2m37s (2,257MB)	2.66x

Should you switch anyway?

Yes. A drop-in, MIT-licensed replacement that measures 2.3-2.7x faster with meaningfully lower memory use is worth installing regardless of whether the exact headline multiplier shows up on your file. I walk through the actual install and the large-v3-vs-turbo model choice in the Whisper setup guide.

Our pick

faster-whisper, every time, for anyone running Whisper outside a notebook. The real 2.3x-plus measured gain and lower memory footprint hold up regardless of the marketing ceiling, which is exactly why it's the backend behind both the Whisper setup guide and the offline Home Assistant voice pipeline on this site.

How I picked these numbers

Every figure in the table above came from SYSTRAN's own published README. I did the division myself instead of repeating their marketing line unchecked. The wider third-party range is cited separately and clearly, since it's a different file length, different hardware set, and a less controlled methodology than the vendor's own side-by-side numbers. Where a row compared faster-whisper to something other than stock openai/whisper (the distil-whisper-vs-Transformers row), I said so instead of letting it blend into the main claim.

Vendors publish their own tables more often than anyone actually reads them. This one had the real number sitting in plain sight the whole time — it just wasn't the number in the first sentence.