MarkTechPost Tests Cactus Whistle's 16.9 MB CPU Speech Model Against Whisper

The model card claims Whistle beats Whisper base on LibriSpeech. The guide compared the systems on seeded test-clean and test-other samples using paired speaker-cluster intervals with Holm correction. It said its own figures were one-run samples from an Apple M4, and that published full-split numbers were quoted rather than measured.

The test also probed a 30-second window in which the pinned engine silently dropped the middle of longer audio, according to the guide. It examined streaming latency, word timestamps, keyword biasing, other languages, noise, embeddings and speech-to-tool calls. The setup installed cactus-needle 3.2.0 with openai-whisper, jiwer, soundfile and soxr, and pinned the Whistle and engine revisions on Hugging Face to make the run reproducible.

MarkTechPost reported that the speed advantage over Whisper narrowed as clip length increased, a caveat to the compact model's CPU appeal. The guide's comparison covered Whisper tiny and base, not only base. It also distinguished greedy decoding from beam search in its accuracy findings.

The guide's headline claim centered on a lead over greedy Whisper base on test-other, while the beam-search comparison showed no significant gap. Its reported limitations included the one-run sampling and the 30-second audio handling, both of which affect how broadly the results should be read.