AI BriefingPickwhistle: 7-Language Speech Recognition in a Single 16.9MB File
Cactus-Compute/whistle
About the project
This is a speech recognition model targeting edge devices such as mobile, wearable, robot, and smart home devices. The entire model is contained in a single 16.9MB file and runs on CPU only, without requiring a GPU. It adopts a lightweight structure that can be executed without external dependencies.
It supports transcription for 7 languages: English, German, French, Spanish, Italian, Dutch, and Polish. It processes 16kHz mono audio up to 30 seconds in length, and also supports word-level timestamps and speech embedding extraction. It provides keyword biasing functionality to prioritize the recognition of proper nouns or product names frequently used by the user.
It shares the same container format and quantization method as the Needle engine from Cactus Compute. If a device is already running Needle, it can load only the Whistle model to perform speech recognition and text generation simultaneously without additional runtimes. It supports C API and WebAssembly, making it easy to integrate into various platforms.
Cactus-Compute/whistle
The original page has no description.
automatic-speech-recognition
This introduction was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report errors, attribution issues, or removal requests via Contact.