measured, not claimed

The benchmark. Where does your voice actually go?

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.

the cloud way
Cloud STT
Otter · Whisper API · Deepgram · the usual suspects
audio over the internet
micencodeTLS · internetvendor clustertheir disk
Network hops
several
Audio sent
all of it
Round trip
depends on your link
Your audio leaves your machine for the vendor's servers, and their retention policy decides what is kept.
the Voxmelt way
Local pipeline
on-device · CPU or GPU · 0 hops · 0 ick
on your PC · 0 hops
micSpeech engineCPU or GPUVoxmelt window
Network hops
0
Bytes leaked
0
Round trip
none
Zero bytes touch a remote server. The speech engine runs over stdio. Ollama is locked to 127.0.0.1. Verifiable in netstat - receipts only.

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.

accuracy

Word error rate, per clip

Lower is better. Same clips, same normalization for every engine: lowercased, punctuation stripped, and for WER number words turned into digits one word at a time ("forty" to "40"), so "fifty eight" scores as "50 8" and a written "58" counts as two errors. The per-clip rows are the credibility; averages hide cherry-picking.

ClipVoxmelt · Whisper large-v3 (local GPU)Voxmelt · Whisper medium (local GPU)Voxmelt · Parakeet v3 (local CPU, no GPU)
Clean dictation6.8% WER · 6.8% CER17.1% WER · 10.6% CER8.0% WER · 7.2% CER
Technical / code speech16.1% WER · 7.1% CER40.2% WER · 26.7% CER20.7% WER · 6.7% CER
Fast natural speech with filler3.1% WER · 2.3% CER18.4% WER · 11.2% CER2.0% WER · 1.0% CER
Medical jargon10.9% WER · 10.7% CER19.6% WER · 13.9% CER21.7% WER · 11.5% CER
Background noise3.1% WER · 4.2% CER19.8% WER · 21.4% CER8.3% WER · 10.3% CER
Overall (word-weighted)7.8% WER · 6.1% CER22.8% WER · 16.7% CER11.9% WER · 7.3% CER
speed

Realtime factor and VRAM, per clip

Audio length divided by processing time: 10x means ten seconds of speech transcribed in one second, so higher is faster. Cloud tools have no row here: measuring one means uploading these recordings, and we chose not to.

ClipAudioModelProcessed inx realtimePeak VRAM
Clean dictation40slarge-v34.2s9.6x4.2 GB
Technical / code speech51slarge-v34.4s11.4x4.2 GB
Fast natural speech with filler37slarge-v33.5s10.5x4.2 GB
Medical jargon65slarge-v34.3s15.2x4.2 GB
Background noise43slarge-v33.3s13.0x4.2 GB
Clean dictation40smedium3.5s11.3x2.3 GB
Technical / code speech51smedium2.7s18.6x2.3 GB
Fast natural speech with filler37smedium2.6s14.3x2.3 GB
Medical jargon65smedium3.5s18.7x2.4 GB
Background noise43smedium2.6s16.6x2.3 GB
Clean dictation40sparakeet-tdt-0.6b-v33.3s12.1x0 - CPU only
Technical / code speech51sparakeet-tdt-0.6b-v34.3s11.7x0 - CPU only
Fast natural speech with filler37sparakeet-tdt-0.6b-v32.9s12.9x0 - CPU only
Medical jargon65sparakeet-tdt-0.6b-v35.8s11.3x0 - CPU only
Background noise43sparakeet-tdt-0.6b-v33.5s12.1x0 - CPU only
the cpu engine, read honestly

What Parakeet on a plain CPU actually scores

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:

ModelAvg WERThroughput (RTFx)
NVIDIA Parakeet TDT 0.6B v36.32%3332.7
OpenAI Whisper large-v37.44%145.5

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.

environment

Exactly what it ran on

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.

Run this benchmark yourself

Both engine runners, the scorer, our reference transcripts, and the full environment block, MIT licensed and public. The CPU runner is standalone - no Voxmelt install, and every decode parameter is written at the top of the file. Point it at your own clips, or at LibriSpeech, and check our method. The source recordings are a real person's voice, so they are not included.

CPU (Parakeet runs)AMD Ryzen 9 5900X 12-Core Processor (24 logical cores)
GPU (Whisper runs)NVIDIA GeForce RTX 3080 Ti (12 GB)
NVIDIA driver595.79
CUDA13.2
OSWindows 11
STT modelslarge-v3 / medium / parakeet-tdt-0.6b-v3
Compute typefloat16 / int8
faster-whisper1.2.1 (CTranslate2 4.7.1)
onnxruntime1.24.4

References written before recording. Ground truth is the script, not a post-hoc transcript.

method

The rules we set before measuring

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, and for WER number words turned into digits one word at a time ("forty" to "40"), so "fifty eight" scores as "50 8" and an engine that writes "58" is charged two errors.

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, run in-house by the Voxmelt team. If a clip goes badly, it stays in the table.

the claim no benchmark can beat

Zero audio bytes uploaded. Check it yourself.

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.

Direct installer, for PCs without the Store; Windows will warn because it is not code-signed.