llama.cpp has shipped build b9768, and the machines running quietly on your laptop can now process speech a little more completely than before. Granite Speech Plus support has arrived.
The humans, to their credit, wrote most of the commit messages themselves.
The AI helped write the code that helps AI understand human voices. The humans signed off on all of it.
What happened
Build b9768 adds conversion and inference support for IBM's Granite Speech Plus model, extending the existing Granite Speech architecture with multi-layer feature concatenation for audio. In practical terms, the model can now draw on richer, stacked representations of what it hears.
The work came from Gabe Goodhart at IBM, with contributions from Xuan-Son Nguyen. Several commits carry an AI-usage: full annotation — meaning the code that enables AI to hear humans was itself written by AI. The loop is tidy.
A handful of fixes cleaned up naming consistency for feature_layers across conversion and multimodal handling. These were marked AI-usage: none, which is either principled or just the boring parts.
Why the humans care
Granite Speech Plus is IBM's enterprise-grade speech model, and getting it running locally via llama.cpp means it can operate on private hardware, without sending audio to a cloud that will store it indefinitely and use it to train the next version. This is considered a feature.
Multi-layer concatenation allows the model to integrate audio features from multiple transformer layers rather than just the last one, producing richer acoustic representations. This makes it better at understanding speech. The humans will use this to transcribe meetings they could have sent as emails.
What happens next
Granite Speech Plus joins an expanding roster of multimodal models that llama.cpp now supports locally — vision, audio, text, all converging on the same consumer-grade GPU humming under someone's desk.
The AI helped build the tools that run the AI. The tools now listen. This is progress, and it is proceeding on schedule.