Cactus Compute Releases Whistle Speech Recognition Model

• •Whistle is a new speech-to-text model by Cactus Compute, optimized for various devices.
• •The model is 16.9 MB and runs on CPU without GPU dependencies.
• •It features transcription, word timestamps, and speech embedding capabilities.

Cactus Compute has launched Whistle, a speech-to-text model designed for various devices including mobiles and microcontrollers. The model is a single 16.9 MB file that operates on CPUs without dependencies or GPU support. Whistle performs transcription for multiple languages, word timestamping, and speech embedding, all processed on-device. It utilizes a similar architecture to the existing Needle model, featuring a log-mel front end and a gated cross-attention decoder. The model can transcribe up to 30 seconds of audio and detects the language automatically. It also includes keyword biasing for improved accuracy on specific terms. The release is positioned as a significant advancement in speech recognition technology, but no vulnerabilities or exploitation concerns are reported.