Efficient Speech Recognition Pipeline for the Automated Medical Reporting

As radiologists worldwide grapple with ever-increasing workloads, the implementation of automated report generation offers a dual advantage: alleviating their workload and enhancing turnaround times. Augmented by Artificial Intelligence (AI)-powered speech recognition, our proposed approach minimizes dictation errors and contributes to precise report generation, thereby reducing the need for post-dictation corrections. While Automatic Speech Recognition (ASR) models have demonstrated exemplary performance in the realm of speech-to-text generation across various applications, the intricate and complex terminology found in medical reporting poses a unique challenge. As a result, ASR models may not consistently deliver optimal results in this domain. To overcome this limitation, we explore the integration of Large Language Models (LLMs) to enhance the performance of ASRs for speech recognition and subsequent report generation. In our study, our objective is to identify the most effective end-to-end Speech_Dictation-to-Text_Report Pipeline. Our dataset comprises 3996 chest dictations in audio format, featuring distinct accents, alongside their corresponding finalized text reports. We initiate our analysis by comparing the performance of ASR models of Whisper, DeepSpeech, Facebook Word2vec, SparkNLP, and Kaldi, as they transcribe the dictated audio files into coherent text. Subsequently, we process the output of the best-performing ASR model through four LLMs of Falcon-7B, Mistral-7B, Qwen-14B, and Zephyr-7B. Among the various model combinations, the pairing of Mistral-7B with Whisper ASR emerges as the top-performing solution, boasting impressive results with a Word Error Rate of 0.031, Sentence Error Rate of 0.019, Match Error Rate of 0.017, and an Average Levenshtein Distance of 0.651.