b9056
No more cut-off audio in local AI transcription—new patch addresses FFT buffer padding
llama.cpp, the widely-used open-source C++ implementation for running LLMs locally, has released version b9056 with a key bug fix for its Whisper integration. The patch addresses an issue where long audio recordings would have their tail end truncated during transcription. The problem occurred because the audio buffer passed to the Fast Fourier Transform (FFT) was not properly padded, causing the last few seconds of speech to be dropped.
The fix ensures the padded buffer is correctly exposed to the FFT, preserving the full audio waveform. This is especially important for users relying on local Whisper models for meeting notes, podcast transcription, or voice interfaces. The release includes binaries for macOS (Apple Silicon and Intel), Linux (x64, arm64, Vulkan, ROCm, SYCL), Windows (x64, CUDA 12/13, Vulkan), and Android. With 109k GitHub stars, llama.cpp continues to be a go-to tool for on-device AI inference.
- Fix for Whisper audio tail truncation by exposing padded buffer to FFT (commit cc97e45)
- Available on macOS, Linux, Windows, Android, and iOS across CPU, CUDA, Vulkan, and ROCm backends
- llama.cpp is the most-starred local LLM runtime with 109k stars and 17.9k forks
Why It Matters
Improves accuracy of local speech-to-text for professionals using Whisper models without cloud dependencies.