ggerganov/llama.cpp Released Version b12238 with Audio Endpoint
Llama.cpp release b12238 introduces an OpenAI-compatible /v1/audio/speech endpoint for text-to-speech functionality alongside various audio handling improvements. It also updates roundings and multipart body parsing for audio helpers to ensure better type consistency. This addition allows developers to drop-in replace OpenAI audio speech calls with local infrastructure running on llama.cpp, facilitating local and offline text-to-speech workflows. It broadens the server's compatibility layer, making it easier to integrate multimodal AI applications locally. The update fixes an issue where the WAV writer truncated data while pcm_f32_to_s16 rounded it, causing a one LSB discrepancy between WAV and PCM formats. Both methods now correctly round, and the conversion handles bytes directly without intermediate copying.
## BACKGROUND
llama.cpp is an open-source software library that performs inference on various large language models like Meta's Llama using the GGML tensor library. It includes a built-in server that mirrors common API formats to simplify integration for developers.