Accuracy and speed for both of Voxmelt's local engines - NVIDIA Parakeet v3 on a plain CPU and Whisper large-v3 on a named consumer GPU - measured on the same real recorded clips with a method anyone can rerun. One sentence, two journeys: hover any lane below to slow the trail and read each hop.
Two engines, one test set, zero uploads. On a stock RTX 3080 Ti, Voxmelt's local Whisper large-v3 averaged 7.8 percent word error rate across five scripted real-world clips at 10 to 15x realtime. And with no GPU at all, NVIDIA Parakeet v3 on a plain CPU scored 11.9 percent on the same clips at 11 to 13x realtime, loaded in 4 seconds, and beat the GPU flagship outright on fast natural speech (2.0 vs 3.1 percent WER) - which is why CPU dictation is now the default.
Measured 2026-07-16. Full per-clip tables below.
Lower is better. Same clips, same normalization for every engine: lowercased, punctuation stripped, spoken numbers folded to digits. The per-clip rows are the credibility; averages hide cherry-picking.
Processing time divided by audio length. Below 1.0 means faster than real time. Cloud tools have no row here because you cannot measure what runs in someone else's building.
Same five real recorded clips, same scorer, same normalization as the GPU rows above. Parakeet v3 (int8 ONNX) on a Ryzen 9 5900X, GPU untouched: 11.9% WER overall vs 7.8% for Whisper large-v3 on the RTX 3080 Ti. On fast natural speech with filler - the clip that looks most like real dictation - Parakeet WINS outright: 2.0% vs 3.1% WER.
Where the overall gap comes from: mostly formatting, not mishearing. Parakeet writes "58", "$1800", and "21st" while the reference scripts spell numbers out, and our strict scorer counts every one of those as an error. The character error rate tells the real story: 7.3% vs 6.1%, nearly even. Its one genuine weak spot in this set is rare drug names in the medical clip.
The comparison that matters for most laptops: on a machine with no NVIDIA card, the alternative to Parakeet is not large-v3 (it needs CUDA) - it is a small Whisper model. Whisper medium scored 22.8% WER on these same clips WITH a GPU. Parakeet gets 11.9% with zero VRAM and a 4-second model load. That is why it is the default engine.
Whisper large-v3 stays in the app as the GPU option: it is still the accuracy king on hard audio (noise, dense jargon) and covers ~100 languages to Parakeet's English + 24 European ones. Voxmelt ships both; the default just no longer requires a graphics card.
third-party context (not our numbers)
On the independent Hugging Face Open ASR Leaderboard - a different, much larger English test set, so these are not comparable to our per-clip numbers above - the same two models rank:
Source: Open ASR Leaderboard results as published in arXiv:2510.06961 (Table 3) and the live leaderboard. Their harness, their hardware, their test sets - cited here so you can sanity-check ours against the wider world.
A benchmark without its environment is an anecdote. Reproduce ours or run your own hardware; Whisper numbers scale with your GPU, Parakeet numbers with your CPU.
References written before recording. Ground truth is the script, not a post-hoc transcript.
Five clips covering the real shapes of dictation: clean speech, technical and code speech, fast speech with natural filler, medical jargon, and background noise. Variety is the point; a benchmark on easy audio only proves you can pick easy audio.
Reference transcripts were written before recording and read aloud. The ground truth is the script, never a cleaned-up transcript of the recording.
Every engine gets the same audio files and the same normalization: lowercase, punctuation stripped, spoken numbers folded to digits. No engine is favored.
Voxmelt is scored on the output of its shipping pipeline, the same refined pass users get, with the same quality gates. We do not benchmark a lab configuration users never see.
Speed counts processing time only, against a warm model. One-time model load is reported separately, the same way cloud tools do not bill you for their server boot.
Caveats stay attached to the numbers: one speaker, English, one hardware configuration, founder-run. If a clip goes badly, it stays in the table.
While every clip above was transcribed, the number of audio bytes that left the machine was zero, because there is nothing to send. Open any network monitor, dictate, and watch. That is the part of this page you do not have to take on faith, and the part no cloud tool can put in their benchmark.