Voice and Speech Experiments

Voice assistant is a convenient name for several systems that can fail independently and then blame one another by producing silence.

My speech experiments keep three basic jobs separate: detecting whether to listen, turning audio into text and generating speech. Model training and evaluation are another layer again. A microphone button is a small interface to a surprisingly large set of dependencies.

The work connects Soma, keyboard dictation and other device interfaces with local compute. Training scripts and manifests live in the model-training repository. Heavier assets and experiments are handled separately, including work on Titan; synthesis has also been assigned to Sisyphus.

Follow the failure to the right component

A correct wake-word trigger does not prove transcription works. An accurate transcript does not prove generated audio reaches a headset. Keeping those stages distinct makes it possible to describe a result without promoting it into a claim about the entire pipeline.

This page is an index for that work rather than a comparison chart with numbers detached from their test conditions. Model choices, transport details and measured results belong in individual notes. I want to be able to trace a statement back to the experiment that supports it, especially when the system looked excellent until someone tried to use it outside the test setup.