During February 2026, Bell Kjtt enhanced the VibeVoice ASR pipeline in the Blaizzy/mlx-audio repository by implementing robust audio preprocessing features using Python and audio processing techniques. Bell introduced resampling and loudness normalization to ensure all input audio is standardized to 24 kHz and -25 dBFS, addressing inconsistencies in input quality and improving transcription accuracy. The solution included sampling_rate-aware APIs for both generation and streaming transcription, enabling seamless handling of varied audio sources. By aligning decoding parameters with the official demo, Bell’s work improved robustness, facilitated easier benchmarking, and provided a more reliable foundation for machine learning-based speech recognition.