The AI Learning to Sing Like a Human

    The new singing bot is called XiaoiceSing. In "singing voice synthesis," the task is to turn a musical score into a song with a voice. The team worked on a way to offer the ability to sing to the company's chatbot, Xiaoice (pronounced Shao-ice). Xiaoice Sing, and the results are impressive. The first research results are available on the website of Microsoft Technology Center Asia.

    XiaoiceSing does this by dividing the worlds into phonemes and then assigning each one a pitch and duration. This can be expressed in the form of a vector that can be' read by a computer, but this process of translation is tricky. Each word is made up of syllables, and they are formed from phonemes in turn. The score might indicate that the entire word is sung for several beats. But XiaoiceSing's problem is dividing those beats between the phonemes.

    XiaoiceSing uses a system called FastSpeech for text-to-speech. It must then be decoded and vocalised, or "vocoded." By associating the spectral characteristics of the human song with the machine-readable score, the machine essentially learns. Then, given a new score the machine has never seen a human-like output can be produced. The results are impressive, with XiaoiceSing repeatedly coming out on top. 

Read More