ggerganov/llama.cpp Released b12061 with PocketTTS Voice Support
Llama.cpp release b12061 introduces support for PocketTTS voices without requiring voice cloning. This allows users to select predefined voice states directly through prompts or command-line options. This update expands the audio generation capabilities of llama.cpp by integrating lightweight text-to-speech features. It makes it easier for developers to build multi-modal applications running efficiently on local hardware. The converter transforms predefined states back into prompt embeddings stored in mmproj files and sets a default voice if none is provided. Additionally, unknown voices now correctly return a 400 bad request error instead of a server failure.
## BACKGROUND
PocketTTS is a lightweight text-to-speech application designed by Kyutai to run efficiently on standard CPUs without relying on heavy GPUs or web APIs. Llama.cpp is a popular open-source C/C++ implementation for running large language models locally with minimal resource overhead.