Speech to text benchmark

6 dictation apps, the same 36 talk clips, one Mac. Updated 19 November 2026.

How we test

Leaderboard

Accuracy and speed for every app and model
Nemotron1
94.2
—3
Parakeet v3
93.5
0.37 s
8.5 s
Parakeet Unified
94.7
0.50 s
—3
Parakeet v3
93.6
0.57 s
17.0 s
Base
71.54
0.66 s
16.1 s
Small
75.44
1.51 s
34.5 s
Turbo
95.0
2.76 s
64.8 s
Turbo
92.1
4.34 s
94.2 s
Turbo
95.0
4.95 s
95.5 s
Turbo2
95.1
15.36 s
69.0 s
  1. Nemotron types as you speak. Its last words appear 0.30 s after you stop.
  2. Buzz loads the app and model for every file, about 13 s.
  3. Handy and Hushpen Nemotron Unified were not run on 30 minute recordings.
  4. Small and Base wrote Welsh or Korean on 6 and 8 of 36 clips (language left on Detect).

Our picks

  1. Near the top, and faster
    Turbo

    95.0 of 100 words right

    0.1 below the top score. Faster than every app that scores as high.

  2. Fast and accurate
    Parakeet v3

    0.48 s until your text appears

    93.6 of 100 words right. The default model.

  3. Accurate and quick
    Parakeet v3 + S1-mini

    94.5 of 100 words right

    With S1-mini cleanup. Text in 0.94 s.

How long you wait after you stop talking

Hushpen only. The full app typing into TextEdit, middle of 3 runs.

Seconds from the end of speech until the text is in the app
Nemotron0.30 s0.29–0.33 s · 3 runs0.79 s0.74–0.89 s · 4 runs1.7 s
Parakeet v30.48 s0.46–0.51 s · 3 runs0.94 s0.94–1.01 s · 3 runs0.6 s
Turbo1.69 s1.68–1.69 s · 3 runs2.17 s2.15–2.22 s · 3 runs0.8 s

S1-mini loads in 1.1 s on first use.

By accent

Words right out of 100, 19 English accents. Higher is better.

Words right out of 100 for each accent
IndianArunachalam Muruganantham, Sugata Mitra8988859089908590
Spanish or Latin AmericanLuis von Ahn, Pia Mancini9087899090908990
ChineseEric X. Li, Yang Lan9293929191949392
AmericanBrené Brown, Simon Sinek9294929394949294
RussianGarry Kasparov9393979390909392
JapaneseShigeru Ban, Takaharu Tezuka9487869595888539
IrishBono, Mary Robinson9494939493939394
South AfricanNeil Turok, Toby Shapshak9590929595939194
GermanRalph Langner, Stefan Sagmeister9595949597949594
BrazilianEduardo Paes, Miguel Nicolelis9695939496949294
ScottishEvelyn Glennie, Gordon Brown9696939696989598
KoreanHyeonseo Lee, Joseph Kim9695979696969595
FrenchEsther Duflo, Matthieu Ricard9797969797979697
NigerianChimamanda Ngozi Adichie, Ngozi Okonjo-Iweala9795959696959596
BritishJamie Oliver, Ken Robinson9799999996999998
AustralianPeter Singer, Rachel Botsman98989710098989898
PakistaniKhalida Brohi, Ziauddin Yousafzai9897989898979897
ArabicRana el Kaliouby, Wael Ghonim9898969898989697
FilipinoMaria Ressa9999999799999799

Best in the rowWorst in the rowHushpen

Two clips per accent. Small and Base are left out. VoiceInk wrote one Japanese-accent clip in Japanese.

What each app says, and what we measured

What each app says, and what we measured
AppWhat the app saysWhat we measured

"Faster than any server can answer."Speed

"Our own purpose-built speech models deliver fast, accurate dictation."Accuracy and speed

"Insanely fast Parakeet"Speed

93.5 of 100 words right. 1 minute in 0.37 s.

"Local AI models that transcribe your voice to text with 99% accuracy, almost instantly"Accuracy and speed

92.1 of 100 words right. 1 minute in 4.34 s.

"Get accurate text transcriptions in seconds (up to ~30x realtime)"Accuracy and throughput

"MacWhisper now supports Parakeet models (up to 300x realtime speeds)"Throughput

95.0 of 100 words right. 1 minute in 4.95 s.

"Fast and accurate, great overall multilingual pick"Speed and accuracy

"Best all-around choice for English dictation"English recommendation

"3.99"Word error rate

"25.3×"Throughput

94.7 of 100 words right. 1 minute in 0.50 s.

Not tested yet

Claims as each app states them, seen on 19 November 2026. Their tests differ from ours.

How we ran the testSame 36 clips for every app, default settings, middle of 3 runs.

Apple M4 Pro Mac, 19 November 2026. Other versions may differ.