Description
Whistle is a speech-to-text model developed by Cactus Compute, available on Hugging Face. It is designed to run on mobiles, wearables, robots, smart home devices, automotive systems, and microcontrollers. The model is compact, with a file size of 16.9 MB, and operates on the same CPU engine as Needle, utilizing the same container and quantization techniques. Whistle performs three primary functions: transcription, word timestamps, and speech embedding.
Read More
The transcription feature processes 16 kHz mono audio, supporting English, German, French, Spanish, Italian, Dutch, and Polish. It can handle audio of any length, performing one-pass transcription up to 30 seconds and windowed transcription beyond that. The model detects the language automatically unless specified, and returns an empty transcript for silence. It can also transcribe live audio streams, committing words agreed upon by two passes.
Word timestamps provide the start, end, and probability for each word, allowing applications to highlight, seek, or cut based on word alignment. The speech embedding feature outputs encoder results for matching and retrieval without decoding transcripts. Whistle includes keyword biasing, favoring names, places, and product words during searches.
Whistle's architecture includes a log-mel front end, a convolutional stem, and an audio encoder, with a Needle-shaped decoder using gated cross attention. It reuses Needle's container, quantization, SIMD kernels, and KV cache, enabling seamless integration with devices already running Needle. The model supports various deployment depths, selectable at load time.
Benchmarks indicate Whistle's performance in terms of word error rate, speed, and precision, compared to other models like Whisper and Moonshine. The Python package and source code are available on GitHub, facilitating easy integration and deployment. Whistle is ideal for developers seeking a robust, multilingual speech-to-text solution for diverse applications.
- CategoryAI Models & LLMs
- Pricing ModelFree
- DeploymentCloud, Self-hosted
- PlatformsWeb
Whistle Highlights
Multilingual transcription
Word timestamps
Speech embedding
Keyword biasing
Log-mel front end
Convolutional stem
Gated cross attention
Seamless integration with Needle
Getting Started with Whistle
Access model: Visit Hugging Face
Authenticate: Set up credentials
Set up environment: Install dependencies
Integrate via API: Use Python package
Optimise: Adjust model parameters
Whistle's Use Cases
- Mobile Transcription
- Wearable Integration
- Smart Home Automation
- Automotive Speech Recognition
- Microcontroller Applications
