Research Publications & White Papers from Sebastien Rousseau

Buy and Download copies of my collection of White Papers and Research Publications on Artificial Intelligence (AI), Quantum Cryptography, and more.

Publications

Purchase PDF - $49.00 Publication can be printed, downloaded but can not be redistributed. Buy Now ❯

Accelerating Real-Time Speech Recognition with OpenAI Whisper and Metal Performance Shaders on macOS

Sebastien Rousseau - 12 March 2024 - English(95 KB - PDF)

Abstract Real-time automatic speech recognition (ASR) has become an increasingly crucial technology in today's world, enabling a wide range of applications such as live transcription, voice assistants, and dictation. However, performing high-quality ASR inference with low latency remains a computationally demanding task, often requiring powerful cloud servers. In this paper, we present a system for real-time speech-to-text transcription that leverages the OpenAI Whisper speech recognition model and accelerates inference on macOS devices using the Metal Performance Shaders (MPS) GPU back-end. By utilising on-device GPU acceleration, our system significantly reduces the latency and computational requirements of real-time speech recognition while maintaining high accuracy. The system is implemented in Python and can easily be integrated into a variety of applications requiring live speech-to-text capabilities. We evaluated the performance of our system on a M1 Max MacBook Pro and demonstrated that it achieves sub-second latency and 8-12 times faster than real-time transcription for typical utterance lengths of 5-10 seconds. Our energy-based voice activity detection approach achieves 94% precision and 96% recall in detecting speech segments. The proposed system showcases the potential of leveraging state-of-the-art ASR models and GPU acceleration for efficient and accurate real-time transcription on edge devices, making it an ideal solution for applications requiring live speech-to-text capabilities on macOS devices.