音樂與人工智慧實驗室
Music and Artificial Intelligence Lab
In our lab, we design and apply machine learning algorithms to music and audio signals. We have four main topics in the lab: AI Listener, AI Composer, AI Performer, and AI DJ. The first one is mostly about content analysis, whereas the latter three are about content generation. Our research are mainly about machine learning but depending on the specific projects our research can also concern with audio signal processing, musicology, human computer interaction, natural language processing, information retrieval, psychology, and edge computing.
AI Listener

We use AI Listener as a blanket term to refer to research topics that are related to music and sound analysis, such as source separation, music transcription, structure analysis, sound event recognition, instrument recognition, and emotion recognition. The goal is to learn computer models that can understand and appreciate music and audio signals in the same way as human beings. Such models can find applications in music retrieval (e.g. similarity search, content based recommendation), music education, and also music generation.

AI Composer

The goal of AI Composer is to build machines that can compose new music. In particular, an AI Composer creates music in the so-called "symbolic" domain, in formats such as melody lines, lead sheets, or multitrack pianorolls (similar to MIDIs). The generation of such musical scores could be either unconditioned (from random seeds) or conditioned (e.g. given a prime melody, given a chord sequence, given the lyrics, given some tags such as genre or emotion tags, or given an image or a video sequence). We hope that our AI Composer models can help musicians or music lovers create music in an interactive way.

AI Performer

Music creation is typically composed of two parts: composing the musical score, and then performing the score (with certain instruments) to make sounds. The musical score mainly specify what notes to be played, but usually not how to play them. Musicians leverages this freedom to interpret the score and add expressiveness in their own ways, to “bring the music to life”. Accordingly, in addition to AI Composer models that deal with symbolic-domain music generation, we need AI Composer models to deal with audio-domain music generation and create musical audio from scores.

AI DJ

Disc Jockeys (DJs) are professional audio engineers whose role is to generate music artwork such as electronic dance music (EDM) and to manipulate musical elements to create music medley, mashup and remix. It is a profession that adds values to music. In addition to generating new music from scratch using AI Composer models, we can also build AI DJ models that generate new music by reusing, recombining and manipulating existing musical pieces, perhaps directly in the audio domain. In addition to considering AI DJ as another way to generate music, we believe AI DJ can also be deployed on smart speakers to deliver personalized music recommendation in a DJ way.

playing band together
2013

group dinner
2014

attending ICASSP'14 at Florence
2014

organizing and attending ISMIR'14 at Taipei
2014

group dinner
2015

teaching music information retrieval at National Tsing Hua University, Taiwan
2016

group lunch
2017

group photo at the main entrance of the CITI building, Academia Sinica
2017

attending ICME'17 at Hong Kong
2017

group dinner
2018

group lunch
2019

attending ISMIR'19 at Delft
2019

group lunch for the team at Taiwan AI Labs
2019

visiting Positive Grid at Taipei
2020

group dinner
2021

group photo with the PI's photo at the 1F the CITI building, Academia Sinica
2021

group lunch for the team at Taiwan AI Labs
2022