Institute of Information Science, Academia Sinica

We may say that this wave of “AI Renaissance” started around 2010 when the so-called deep learning technology was applied to automatic speech recognition (ASR) and for the first time achieved an apparent success in a serious application of sufficient scale. The performance lift for ASR obviously continued in the following years and naturally leads to a question: “Is the ASR a solved problem?” The answer may be “yes” for phonetic classification or for read speech recognition. However, read speech is only a special case in our daily life and it’s far from understanding an utterance if you can only distinguish phonetic classes. There used to be three major challenges in ASR proposed in 1980s. By contrast, a new “three major challenges” will be proposed in this talk. Above all, issues of spontaneous speech have been collected and organized to show the numerous obstacles in the way to the ultimate solution of the ASR problem. Various examples of real cases conducted in ITRI will be shared and demonstrated during the talk.