- Dictation
- Engineering
- Performance
Whisper vs Parakeet: on‑device speech to text
Both models run on a Mac without the cloud. On the same clips they match for accuracy — they separate on speed, memory, and languages. Measured numbers inside.
If you’re picking an engine for dictation on a Mac, the shortlist is mostly two names: OpenAI’s Whisper and NVIDIA’s Parakeet. Both run entirely on your machine, both are open enough to ship, and — the part that surprised us — on the same audio they’re equally accurate. What separates them is speed, memory, and how many languages they speak. This post has the measured numbers, from the bench we built when choosing the engine for Snapbak.
Two models, two philosophies
Whisper is OpenAI’s general‑purpose speech model: MIT‑licensed, covering 99 languages including Hindi, with a turbo variant tuned for speed. Parakeet is NVIDIA’s family of models built on the TDT architecture — fast by design. The version Snapbak ships, Parakeet TDT 0.6B v2, is English‑only; NVIDIA’s newer v3 extends to 25 European languages.
Accuracy: closer than the leaderboards
Published benchmarks put Whisper large‑v3‑turbo at a mean 7.83% word error rate and Parakeet TDT 0.6B v3 at 6.34% across their respective suites — figures from the published model cards and the Open ASR Leaderboard, and worth reading carefully, because they use different test sets and shouldn’t be compared directly.
So we ran our own: both engines, the same 60 LibriSpeech test‑clean clips, same machine. Both scored 2.3% word error rate. On clear speech, accuracy is a tie — choose on everything else.
Speed: where they separate
Measured on an M4 with 16 GB of memory, macOS 15.5, models loaded:
| Utterance | Whisper large‑v3‑turbo | Parakeet TDT 0.6B v2 |
|---|---|---|
| 3 seconds of speech | 0.68 s | 0.045 s |
| 10 seconds of speech | 0.97 s | 0.058 s |
| 30 seconds of speech | 2.51 s | 0.166 s |
The Parakeet figure with a 200‑term custom dictionary loaded was 0.082 seconds for ten seconds of speech — still an order of magnitude faster. This is why dictation built on Parakeet feels instant: the model is never the wait. End to end, from releasing the key to text inserted in the app you were typing in, the median we measure is 0.44 seconds.
Memory
The helper process holding the model peaked at 629 MB with Parakeet against 852 MB with Whisper turbo — 228 MB less, and the weights are essentially all of it. After the helper unloads, Snapbak idles at a measured 25 MB. Neither model ships inside the installer; both download on demand (Parakeet v2 plus its CTC variant is about a 540 MB first‑run download).
The real trade: languages
Here is the whole trade in one line: Whisper covers 99 languages including Hindi; Parakeet v2 is English‑only. That is why Snapbak dictates in English and nothing else — Parakeet’s speed and memory won the job, and the language coverage was the price. If you need to dictate in Hindi or another of Whisper’s languages, a Whisper‑based tool is the right call today, and we’d rather say that than pretend otherwise.
What to check when you pick a dictation app
Whatever the engine under the hood:
- Which model it runs, and where — on the Neural Engine, the GPU, or a server. “On‑device” claims deserve the follow‑up question.
- When the model loads — Snapbak warms the helper at key‑down and quits it 60 seconds after the last dictation, so the memory is spent only while you’re talking.
- Whether accuracy figures are measured or borrowed — published leaderboards and an app’s real behavior on your vocabulary are different claims.
Try it
Snapbak dictates English, on‑device, into any Mac app — the engineering numbers are documented, the app options are compared, and the workflow is the part you’ll actually feel. Download Snapbak free for the first 100 users.