They are applied both to the acoustic model – translating sound to words – and to the language model which deals with how words statistically tend to appear together. However, the DNN’s are operating continuously even while the user is using the software. In this version 15 of DNS, DeepLearning algorithms have been trained in advance on GPU’s (enabling parallelisation) using large datasets of speech (big data). With Dragon NaturallySpeaking 15 we should expect to see even more improvements.ĭeep learning algorithms should imply that the speech recognition is even more adaptive to the individual user, even more robust to noise, has a higher ability to learn from mistakes, dialectic pronunciations, context of words, individual writing style et cetera.
A new language model called BestMatch VI was introduced, significantly improving accuracy and speed. This was even recommended at some forums. It was possible to start using the software without training at all. The transition from version 12 to version 13 was impressive both in accuracy, robustness to noise and the significantly lower time it took to train the software before it could be used. Because DeepLearning – using Deep Neural Networks (DNN) – had its breakthrough in 2009, my misunderstood assumption was that DeepLearning already played a part in the software by the release of Dragon NaturallySpeaking 13 (in 2014). Hidden Markov models are statistical methods which has been around since the 1960s and one of the first applications of HMM was actually towards speech recognition in the 1970s. The machine learning algorithms used in speech recognition since then has been based on Hidden Markov Models (HMM). Even though they are impressive, given a mobile platform, they are far out of reach as an efficient work tool compared to the professional speech recognition solutions offered by Dragon NaturallySpeaking on a PC/Mac.ĭragon NaturallySpeaking was first introduced in 1997.
Deep learning for speech recognition has been used by Google and Baidu for a few years already, but they are still limited to the mobile market. This version introduces something completely new on PC/Mac-based speech recognition solutions.: A deep learning speech engine. Nuance Communications released (September 1, 2016) version 15 of Dragon NaturallySpeaking.
0 kommentar(er)
