Which on-device speech recognition is fastest on a Mac?

Apple's SpeechTranscriber is the fastest on-device engine on a Mac in our test, then Parakeet TDT v3, then Whisper large-v3-turbo. On English recordings Apple transcribed 34 times faster than real time ¹, Parakeet 19 times ² and Whisper 12 times ³. "Times faster than real time" (also called RTFx) means seconds of audio transcribed per second of compute: a value of 20 turns one minute of speech into text in three seconds.

All three speeds are far above what dictation needs. A person speaks in real time, so any engine above 1 times real time can keep up with a live speaker in principle. The differences matter for transcribing recordings: a one-hour file takes minutes on a fast engine and a few minutes more on a slower one. Each engine was timed on the same Mac (an M2 Pro) with the same 100 FLEURS recordings per language. Our benchmark dataset has the accuracy side of the same runs.

Apple is about two to four times faster than the open engines

Speed in times faster than real time, same Mac, higher is faster:

Apple's speed varied by language, from 34 in English to 75 in Italian ¹³. We did not test why speed varies by language. Apple's Portuguese run is left out: it reported 51177 times real time ¹⁶ because it produced no usable text (a word error rate of 100.00 percent ¹⁷). A speed without a transcript is not a speed.

Parakeet is faster than Whisper in every language we timed

Parakeet ran between 18 times real time in Bulgarian ¹⁸ and 22 in Spanish ¹¹. Whisper ran between 11 in Greek ¹⁹ and 17 in Italian ¹⁵. The two ranges do not overlap, so Parakeet is faster in practice, by roughly one and a half times. All rows from this machine:

Showing 8 of 42 rows. See every row in the dataset.

On a data-center GPU the gap is much larger

On a GPU the picture is similar but more extreme. On the Open ASR Leaderboard's English tests, Parakeet TDT 0.6B v3 had an RTFx of 3,330 ²⁶, while Whisper Large v3 Turbo had 200 ²⁷. Those figures come from GPU hardware and large batches, so they say nothing about your Mac. They do agree on the order: Parakeet is built for speed. The leaderboard paper explains why: token-and-duration transducer decoders offer superior speed ²⁸. The speed dataset lists every figure with its hardware.

What this means for dictation

• Live dictation: speed is not the limit. Any of the three keeps up with speech. Latency to the first word, which we did not measure, matters more.
• Transcribing recordings on a Mac: Apple's engine is fastest where it supports your language and works. Parakeet is the faster of the open engines.
• Accuracy first: the fastest engine is not the most accurate. Apple was behind in word error rate in our test, as shown in the Apple against Whisper comparison. The Whisper against Parakeet comparison has the accuracy trade-offs of the two open engines.

What the data does not show

• One Mac. All three engines ran on a Mac with an M2 Pro. Newer chips are faster, and the order could change with a different software build.
• Batch transcription, not live streaming. We timed whole short recordings, not delay while someone speaks.
• Short clips. FLEURS recordings are single sentences. Long files may run at different speeds.
• Whisper model size. We timed large-v3-turbo only. Smaller Whisper models are faster and less accurate.

Related reading

See the guide to on-device speech recognition and our overview of speech recognition. Related questions: can Whisper transcribe in real time, can Whisper run without a GPU and what NVIDIA Parakeet is.