Whisper vs. Parakeet: real benchmark numbers
Most speech-engine comparisons on the web quote unsourced numbers. This page shows exactly how each number was produced and ships the script that produced it — both engines measured side by side, on the same machine, under the same conditions.
Get WhisPaste, free and open sourceMethodology
The table below shows how fast and how accurate each engine actually is, measured under identical conditions on the same machine — no marketing rounding. Real-Time Factor (RTF) is processing time divided by audio duration — lower is faster, 1.0× means exactly real-time. Word Error Rate (WER) is the standard word-level edit distance (substitutions + deletions + insertions) divided by the reference word count, micro-averaged across every utterance in the test set (summing edits and reference words first, not averaging per-utterance WER, which would over-weight short sentences). The test set is 20 utterances (about 94 seconds total) from LibriSpeech dev-clean (CC BY 4.0, Panayotov et al. 2015) — read speech from public-domain audiobooks with known ground-truth transcripts. That is still a small, clean-speech sample — real-world audio with accents, background noise or fast speech will score worse. Treat every row here as a floor, not a guarantee. whisper.cpp is listed twice on purpose: once with its GPU backend (Metal) enabled, and once with it explicitly disabled (-ng/--no-gpu), so the CPU-only row is a fair, same-hardware comparison against Parakeet, which has no GPU backend at all.
| Hardware class | Engine | Model | Speed | Accuracy | |
|---|---|---|---|---|---|
| Apple Silicon (Metal) | whisper.cpp | small (Compact) | 9× faster than real-time RTF 0.1173× | 96.6% of words correct WER 3.4 % | Measured |
| Apple Silicon (CPU) | whisper.cpp | small (Compact) | 3× faster than real-time RTF 0.3068× | 96.6% of words correct WER 3.4 % | Measured |
| Apple Silicon (CPU) | Parakeet (sherpa-onnx) | parakeet-tdt-0.6b-v3 | 35× faster than real-time RTF 0.0288× | 98.1% of words correct WER 1.9 % | Measured |
The difference at a glance
Both engines on CPU only, same machine — the fair baseline, no GPU advantage on either side.
How these numbers were measured
These numbers come from one Apple Silicon machine, each engine measured once under the same conditions. Other hardware — different chips, thermal state, background load — will get different numbers. The scripts below show exactly what was measured and how, in case you want to check the details.
View the whisper.cpp benchmark script on GitHub · View the Parakeet benchmark script on GitHub