Research And Implementation Of Environment Adaptation Technology In Radiology Speech Recognition

In recent years, with the quality improvement of radiology medical equipment in medium-sized hospitals, the daily number of diagnostic is rising rapidly, and also the workload of the film reading doctors. Along with the image displayers replacing the film gradually, the diagnostic report writing methods and tools have been moving to the digital stage, followed by the issue how to efficiently use report writing tools. Before in some hospitals the voice conversion workers turn radiology diagnostic doctors" oral reports into the final report, meanwhile, it is a time-consuming and high labor method. Since the voice recognition technology continues to mature, it has been applied to the construction of medical information. The radiologists can operate the speech recognition software and generate diagnostic reports directly to control time and quality, no longer considering the transcription. The technology has been successfully used in the radiology, pathology, emergency room and other department of US leaded Western countries’hospitals, greatly improves work efficiency and reduces the hospital’s day-to-day operating costs. However, in China the continuous Chinese speech recognition technology has not been successfully used to write diagnostic reports. One of the main reasons is that the recognition rate and recognition speed was unsatisfied because of the high-density of patient population. The noisy environment not matching with the lab training environment conducted a badly impact on the effects of recognition system. How to improve the environmental robustness of continuous speech recognition systems becomes a key factor to the technology popularity in the domestic medical field.Based on Sphinx, the source open speech recognition engine, we designed a continuous speech recognition system specifically applied to radiology diagnostic imaging reports. Based on the system we analyzed the environment adaptive algorithms in robust speech recognition, and designed the complete adaptive process. The experiment results demonstrated that the proposed environment adaptive method provides good performance of the system in a noisy environment.Firstly, the paper discusses the key technologies of speech recognition, mainly including the core idea of the HMM algorithm, as well as the acoustic model HMM-based modeling, which lay the theoretical foundation for building a radiology speech recognition system and adaptive technology. Secondly, we analyzed the noise impact on the voice recognition system and common voice anti-noise technology, then proposed a new environment model adaptive algorithm, based on a synthetic adaptation method which introduces a simplified maximum likelihood linear regression (MLLR) module to the incremental maximum a posteriori (MAP) processing; Finally, with the open-source speech recognition engine, we build up the radiology speech recognition system, designed the noise environment adaptive process and testing process. With brain and lung X-ray radiographic report entry corpus, we recorded the training and testing voice data, and then training acoustic models, which further adapted with noisy speech data set. Finally, experimental results are analyzed. The experimental data proved the algorithm improved the performance of speech recognition systems in noisy environments.