☰
Other toolsWhistleResearch & Data2026-10-15

Whistle vs Whisper on the JFK moon speech

Our own text card comparing Whistle (Cactus) and Whisper base.en on JFK's 'We choose to go to the Moon' speech at Rice University, 1962, full 17 min 40 s, on an 8-vCPU Intel Xeon with no GPU. A table lists model file 16.9 MB and 145 MB, license Apache-2.0 and MIT, run with cactus-needle 3.2.3 and faster-whisper 1.2.1 in fp32, and words out 2,216 and 2,224. For 'distinguished guests' Whistle wrote 'stingwisch guests' and Whisper was correct; for 'But why, some say, the moon?' Whistle wrote 'I quiet some say the moon' and Whisper was correct; for 'moon in this decade' Whistle wrote 'this detain' and Whisper 'dislocate'; for 'President Pitzer' Whistle wrote 'Timmermans' and Whisper 'Pipsa'. A note below says Whistle is about a ninth the size and Whisper base got more of the famous lines right.
Our own card. Audio: public domain (US government work), from the JFK Library via Wikimedia Commons.

We ran two free speech-to-text models on JFK's full "We choose to go to the Moon" speech at Rice University (17 min 40 s of public-domain audio from the JFK Library, via Wikimedia Commons), on an 8-vCPU Intel Xeon cloud machine with no GPU, both on default settings. Whisper got the words of "distinguished guests" and "But why, some say, the moon?" right, where Whistle wrote "stingwisch guests" and "I quiet some say the moon". Both garbled "We choose to go to the moon in this decade" and missed the name Pitzer. Run times swung too much on our shared machine to compare speed.

Try it yourself

What you get

A plain text transcript of a long recording from a free model that runs on your own computer, with no GPU.

Tools

  • Whistle from Cactus: a single 16.9 MB file (Apache-2.0), pip install cactus-needle
  • Whisper base.en: a 145 MB file (MIT), run with faster-whisper in fp32
  • Any recent CPU; we used an 8-vCPU Intel Xeon, no GPU, default settings

Steps

  1. Install Whistle with pip install cactus-needle (we used 3.2.3) and Whisper base.en with pip install faster-whisper (we used 1.2.1).
  2. Convert your recording to a 16 kHz mono WAV file.
  3. Run both on default settings. For Whisper base.en we used fp32, because int8 gave garbled text on our machine.
  4. Read both transcripts against the official text. Names and the line "We choose to go to the moon in this decade" came out wrong in both, so check those by hand.

Watch out: Whistle is Apache-2.0 and Whisper base.en is MIT. The audio is public domain (US government work), from the JFK Library via Wikimedia Commons. One recording only. Run times swung too much on our shared machine to compare speed.

Proof / source

Tested by AgentPlays Our own test on Oct 15, 2026, on an 8-vCPU Intel Xeon cloud machine with no GPU, both models on default settings: the audio, both transcripts and their word counts (2,216 for Whistle, 2,224 for Whisper).

Read the original source ↗ The speech audio ↗

Added · How we verify · Report an error

Related