Improving speech recognition using dynamic multi-pipeline API

Speech recognition technology has been helping human more comfortable to input to computer, however, there is a concern about its performance in terms of precision and error rate in noise environment. Moreover, most of speech recognition APIs are model audio dependent. Although they support the model, the error rate is not decreased. This causes a faulty sound recognition and return incorrect results. This research aims to increase the performance of the speech recognition of non-native speaker. The proposed technique, Dynamic Multi-Pipeline API (DMP), integrated two-tier speech recognition using a combination of Microsoft API, English voice model, and Google API, a Thai voice model, to increase speech recognition performance. The experiment also conducted in different noise scenarios which are 46 dB and 70 dB respectively. The results showed the comparison of performance of the proposed technique with famous speech recognition APIs, the Microsoft API and Google API. It can be seen that the Dynamic MultiPipeline API (DMP) performed better by providing less error rate at 6% and 14% in 46 dB and 70 dB noise environment which are better than compared techniques.