On this page
Most Whisper speed claims come without a method, so they die the moment someone re-measures. This page holds whisper speed benchmarks: measured on real hardware, this month, with the audio corpus, the software commit and the exact commands stated so anyone can rerun them. The short version: on a used RTX 3090 Ti, large-v3-turbo, the large Whisper variant built for fast transcription, processes speech at about 40 times real time, which is an hour of audio in about 90 seconds. On CPU the picture inverts and the model you pick sets the speed, from 15x down to 5x across the ladder on server cores, with laptops far below. Two findings surprised us and change what you should run. On a GPU, smaller models are barely faster than the large one, so the common advice to trade accuracy for speed only applies when you have no GPU. And quantization, widely described as a speed lever, changed GPU speed by a few percent at most; it buys memory, not time.
The numbers
One corpus, one machine, three runs per configuration, medians reported. The audio is 10 minutes 4 seconds of continuous clean English speech; times below are scaled to one hour of audio for readability. Speed is expressed as a multiple of real time: 40x means one hour of recording processed in 90 seconds.
The GPU column is an NVIDIA RTX 3090 Ti, a consumer card two generations old. The CPU column is 32 cores of an AMD EPYC 7702P, Zen 2 server cores; a modern desktop CPU with fewer, faster cores will land in the same region, and a laptop below it.
| Model | GPU speed | GPU, 1 hour of audio | CPU speed (32 cores) | CPU, 1 hour of audio |
|---|---|---|---|---|
| large-v3-turbo (q8_0) | 40x | 1 min 30 s | 5.1x | 12 min |
| large-v3-turbo (f16) | 40x | 1 min 29 s | not run | not run |
| large-v3-turbo (q5_0) | 38x | 1 min 36 s | not run | not run |
| small.en | 40x | 1 min 30 s | 8.2x | 7 min 18 s |
| base.en | 49x | 1 min 14 s | 15x | 4 min |
How we measured, exactly
The corpus is 604.5 seconds of LibriSpeech dev-clean, the first 81 files in speaker-ID order concatenated and resampled to 16 kHz mono PCM. LibriSpeech is public domain audiobook speech under CC BY 4.0, so anyone can rebuild the identical file from the published list. Continuous read speech is close to a best case: no silence for voice activity detection to skip, no crosstalk, no music.
The software is whisper.cpp at commit 6fc7c33 (July 2026), the CUDA build for GPU runs and the plain build for CPU runs, with default decode settings, which currently mean beam search with 5 beams. CPU runs used 32 threads pinned to 32 otherwise-idle physical cores. Each configuration ran three times and the median is reported. The quoted speed is whisper.cpp's own processing time, which excludes loading the model; loading adds about one second for the large model on GPU and under half a second elsewhere, which matters for a ten-second clip and disappears against an hour.
One methodological note from getting this wrong ourselves: the machine also serves production, so every run was checked against an empty job queue, and the first pass of CPU numbers was thrown away because a background embedding model woke up mid-run on the same cores. Speed numbers taken on a busy machine are quietly poisoned; if you benchmark, watch what else is running.
Want to try it now? Upload a file to FastScribe. One file a week is free, no signup.
Why the model barely matters on GPU and decides everything on CPU
The GPU rows cluster near 40x regardless of model size: large-v3-turbo and small.en tie, and even base.en, a small fraction of the compute, only reaches 49x. On a card this fast the time is going into the fixed machinery of transcription, decoding steps, memory movement and the beam search bookkeeping, rather than into the model weights. The practical consequence is blunt: on a GPU there is no speed reason to accept a weaker model. Run the strongest one.
Quantization tells the same story from another angle. f16, q8_0 and q5_0 versions of the same model land within a few percent of each other, with the smallest quantization slightly slower, not faster. Quantize to fit a model into VRAM, not to speed it up.
On CPU the ordering flips and spreads. base.en runs at 15x real time, small.en at 8.2x, and the large turbo model at 5.1x, an hour of audio in about 12 minutes. Note what that means on strong hardware: even the large model stays entirely usable on 32 server cores. The spread matters most below this class of machine, because a four-core laptop lands at a fraction of these figures and the model choice becomes the difference between waiting minutes and waiting hours. If your machine only has a CPU and the transcripts matter, our guide to running Whisper yourself covers the setup and what to do when Whisper is too slow on a laptop covers the escape routes.
What this means for a real recording
These are clean-speech numbers, so treat them as the fast end of the range. A real interview carries silence, crosstalk and room noise; voice activity detection skips the silent stretches, which can make sparse recordings finish faster than these figures while messy dense ones run slightly slower. The processing speed also says nothing about the review pass afterwards, which for most transcripts is the real time cost; how long transcription actually takes budgets that honestly, and the free transcription time calculator turns your own file length into an estimate.
For scale: at 40x real time, a one-hour meeting is about 90 seconds of processing, a full eight-hour deposition day is about 12 minutes, and a ten-minute voice memo is 15 seconds plus a second of model loading. Typing the same hour by hand takes four to six hours, which is the comparison that matters.
Quoting these numbers
Cite them as: large-v3-turbo transcribes at about 40x real time on an RTX 3090 Ti, roughly an hour of audio in 90 seconds (FastScribe benchmark, August 2026). Link this page rather than copying the table, because we re-measure when whisper.cpp or the hardware changes and the page moves with the software; a copied table is stale the day whisper.cpp ships a faster decoder.
A provenance note, because it is the point of the page: we previously quoted 55x from an early measurement whose exact conditions were not recorded, and it did not reproduce under the controlled settings above. That number is retired everywhere. A speed claim without a written method is a rumor with digits in it, ours included; this page exists so the number you repeat has a method attached.
Key takeaways
- large-v3-turbo runs at about 40x real time on an RTX 3090 Ti: an hour of audio processed in roughly 90 seconds.
- On a GPU the model barely matters: small.en ties the large model and base.en gains about a fifth. There is no speed reason to run a weaker model on a GPU.
- Quantization is a memory lever, not a speed lever: f16, q8_0 and q5_0 tie within a few percent on GPU.
- On CPU the model sets the speed: base.en 15x, small.en 8.2x, the large turbo model 5.1x on 32 server cores. Scale all three down hard for a laptop.
- Model load is about a second: irrelevant for an hour, dominant for a ten-second clip.
- The corpus, commit and commands are published; distrust any speed claim, including this one, that does not say how it was measured.
Where FastScribe fits
These are FastScribe's own production numbers: the site transcribes on this exact model, binary and GPU, so the table above is what actually runs when you upload a file, plus voice activity detection and language detection, and queue wait when other jobs are ahead of yours. Your first file needs no account, up to 10 minutes and 50 MB, one per rolling week, with speaker labels included; a free account gives 5 files a day at up to 30 minutes and 100 MB each; Pro is $12 a month for files up to 5 hours and 2 GB with labels on every file, batch uploads and DOCX export. Audio never reaches a third-party AI service and is deleted the moment the transcript is ready. The speed you experience end to end is upload plus queue plus the processing measured here; on a quiet queue a one-hour file returns in roughly two minutes, and no automated tool's speed removes the review pass a transcript you will publish still needs.
Frequently asked questions
How fast is Whisper on a GPU?
On an RTX 3090 Ti we measured large-v3-turbo at about 40x real time with default beam search: an hour of clean speech in roughly 90 seconds of processing, plus about a second to load the model. Smaller models were barely faster, because at this speed the fixed decoding machinery dominates, not the model weights.
Do I need a GPU to run Whisper?
No, but the model you can afford changes. On 32 server-class CPU cores, base.en ran at 15x real time, small.en at 8.2x, and even the large turbo model held 5.1x, an hour of audio in about 12 minutes. On a laptop CPU expect a fraction of these figures, and see our page on Whisper feeling too slow for the practical options.
Does quantization make Whisper faster?
Not on a GPU in our measurements: f16, q8_0 and q5_0 versions of large-v3-turbo landed within a few percent of each other, and the smallest was slightly slower. Quantization is for fitting a model into limited memory. If you saw a speed gain after quantizing, the likelier cause is that the smaller file stopped memory from thrashing.
Why is my Whisper slower than these numbers?
Usually one of four things: a Python implementation rather than whisper.cpp or another optimized runtime, greedy versus beam settings that differ from the defaults here, a busy machine, or thermals on a laptop. The busy-machine one is sneaky: our own first CPU pass was invalidated by an embedding model waking up on the same cores mid-benchmark. Measure with everything else quiet.
How long will my one-hour file take?
At the measured 40x, about 90 seconds of processing on our hardware, plus upload time and any queue wait. By hand the same hour is four to six hours of typing. The transcription time calculator on this site turns your file's real duration into an estimate for typing it, reviewing it, or running it through software.
Try FastScribe on your own recording
One free transcription a week, no signup. Audio is deleted the moment your transcript is ready.
or drop it here · audio or video · one free transcription a week, no signup