Blog
Thoughts on speech, AI, and building in the open.
The Hidden Engineering Behind Real-Time Dictation
Offline speech-to-text needs to be fast - under 300ms total latency. Here is how floure optimizes Whisper inference, audio buffering, and text output to make real-time dictation feel instant on consumer hardware.
Inside an Adaptive Voice Activity Detector
Voice Activity Detection is the unsung hero of speech recognition. Here is how floure's adaptive VAD knows when you are speaking - even in a noisy room - without you pressing a button.
Embeddings Explained Without the Math
Word embeddings are how AI understands human language. No formulas, no PhD required - just a clear explanation of how speech-to-text models like Whisper turn your voice into numbers a computer can work with.
Building AI Software That Thinks in Pipelines
Most AI tools are black boxes. floure is built as a transparent AI pipeline - from audio capture to text output. Here is how pipelined architecture makes local speech-to-text faster, more reliable, and easier to debug.
