Whisper
4 storiesWhisper is OpenAI's family of open speech recognition models, widely used as a baseline for transcription. Recent coverage treats it mainly as a comparison point: Cactus released Whistle, a 16.9MB model scoring 4.31 WER on LibriSpeech test-clean versus 4.9 for Whisper base at 145.3MB, and Lokutor's Oído beat Whisper tiny.en on an ESP32-S3. A macOS on-device search tool also uses Whisper for dialogue search.
Related topics
A small study used 40 synthetic clips with background chatter to test whether speech models could serve as an audio CAPTCHA. Humans averaged 0.71 word accuracy and a 77% pass rate, versus 0.54 and 55% for instructed Gemini 3.8 Flash and 0.55 for Qwen3-ASR—too narrow a gap for a reliable test.
Cactus Compute released Whistle, an open-weights ASR model with 55M params (36M active) that compresses to a 16.9MB CQ2bit file and supports seven languages across 17 platforms. It scores 4.31 WER on LibriSpeech test-clean versus 4.9 for Whisper base at 145.3MB, targeting budget phones and wearables.
A macOS AI search tool enables semantic search across every photo and frame of video in any folder, with OCR text and Whisper dialogue search, running inference fully on-device with no accounts or cloud uploads.
The Lokutor team released Oído, an open-source speech recognition model using a 13M-parameter int8 NVIDIA Conformer-CTC Small that runs on an ESP32-S3 with 8MB PSRAM, no GPU or NPU. It achieves LibriSpeech WER of 3.7/8.2 versus 6.3/15.9 for Whisper tiny.en.
That is everything