Fast, accurate speech to text on the Neural Engine

Tony runs Parakeet Ultra on the Neural Engine of your Mac. A sentence takes about 50 ms, and its English word error rate is 5.8%.

Fast

You speak for	Tony transcribes it in
5 seconds	49 ms
13 seconds	70 ms
26 seconds	147 ms

Parakeet Ultra on the Neural Engine of a MacBook Pro with M5 Pro, median of 5 runs. The clips were read by a macOS voice, and every one came out word for word.

Your first word counts. The audio engine stays prepared between dictations, so capture starts on key down, not a beat later.
Warm from the first press. The model loads at launch and pays its Neural Engine setup up front, so the first dictation of the day is as quick as the hundredth.
Talk for five minutes, wait for none. Tony transcribes in the background at each pause while you keep talking, so letting go is as instant after a long ramble as after one sentence.
Back to back. Press the key again while your last dictation is still landing: it lands first, and the new one follows it.
Idle is free. Between dictations the mic is closed, so no orange dot, and the CPU rests.

The menu bar panel shows your own number: Key up to text, how long your text takes to land after you let go, on a typical dictation.

Accurate

Tony runs Parakeet Ultra: NVIDIA’s Parakeet TDT v3, post-trained by Moondream to make fewer errors on every benchmark they report.

Model	English word error rate	Parameters
Whisper large-v3	7.44%	1.55 B
Parakeet TDT v3	6.26%	0.6 B
Parakeet Ultra, in Tony	5.80%	0.6 B

Lower is better. Averages over the Open ASR Leaderboard English sets: Parakeet from Moondream’s model card, Whisper from the leaderboard paper.

Across 25 languages (FLEURS), Ultra cuts the error rate from 11.6% to 9.6%, and in background noise from 6.7% to 5.8%.

What makes the text read right

Punctuation and capitals included. The model writes them, so the text reads like you typed it.
No “um”. Filler words come out, and the sentence around them keeps its commas and capitals.
Fits the cursor. Tony reads the text before the cursor: a space after a word, none after an opening bracket, and no capital in the middle of a sentence.
Silence types nothing. Voice activity detection trims the quiet, so a cough or a quiet room never turns into a phantom sentence.
Switch languages mid-dictation. Where the model loses track of a switch, Tony transcribes each language on its own. See Languages.
Your words, spelled your way. Names and jargon the model doesn’t know. See Your words.

Under the hood

Parakeet Ultra, 0.6 B parameters, writes the text.
Silero VAD finds where you spoke.
Parakeet CTC 110M listens for your words.

All three run through Core ML with FluidAudio, on your Mac, with no server in the loop. See Privacy and offline use.