Hi, I'm Joren. Welcome to my website. I'm a researcher in the field of Music Informatics, Music Information Retrieval, and Computational Ethnomusicology. Here you can find a record of my research and other projects I have been working on. Learn more »
From 11-16 October 2020 the latest instalment of the ISMIR conference series was held. Due to the pandemic, the 21st ISMIR conference was the first virtual one. As usual, participants and presenters from around the world joined the conference. For the first time, however, not all participants synchronised their circadian rhythm. By repeating most events with 12h in between, the organisers managed to put together a schedule befitting nearly all participants.
The virtual format had some clear advantages: travel was not needed, so (environmental) cost was low. Attendance fees were lower than usual since no spaces or catering was needed. This democratised the conference experience and attendance reached a record high.
Form the 1st of October 2020 I will start on a new research project. The BOF fund of Ghent University is kind enough to sponsor the project for three years. The abstract is as follows:
Music is present in every culture in the world. We as a species seem to have an urge to make music. While the diversity of music cultures around the world is phenomenal, they do seem to have patterns in common. Especially for pitch, one of the fundamental building blocks of music, there are strong reasons to believe that there are commonalities amongst cultures on how pitch is organised A better insight in these common patterns may help to answer questions on the definition, origins and evolution of music.
Common patterns in pitch organisation can be studied from two perspectives. Firstly, the perspective of how humans perceive and make music can be gained from systematic, experimental work. Over the years this has yielded insights in which pitch organisations might be most fit for our perceptual, neurophysiological system. Secondly, these patterns can be observed directly in large-scale, corpus-based, cross-cultural studies which has a potential that is not exploited as of yet.
During this fellowship a large-scale global corpus with field recordings will be compiled in collaboration. Music Information Retrieval techniques will be employed to describe how pitch is organised in the corpus. More specifically, it will support claims on the use of discrete pitches, octave equivalence, the number of pitch classes in use and the pitch interval structures. The uncovered fundamental properties of pitch will be confronted with findings from experimental work.
Recently I presented the outline of the project with the following slides:
A good year ago I was asked to develop audio recognition technology for an e-costume. The idea was that lights in the costume would follow a sequence synchronised to a certain song. Only a single song should trigger the lights, all other music should be ignored. Recognition of music and synchronisation is typically done using audio fingerprinting techniques. The challenge was that the recognition needed to run on a cheap, battery-powered microcontroller with limited CPU and memory. I delivered a prototype but eventually a cheap, battle-tested, off-the-shelf, IP-cleared, alternative was found.
The prototype gathered dust for a while but the idea stuck in my head. With my daughters fourth birthday approaching during the lockdown, I decided to turn the prototype into an over-engineered birthday gift and let an ‘Elsa-dress’ react to ‘Let It Go’ from the Frozen soundtrack. With the prototype as a starting point, I ordered an RGB-LED-strip, a beefy Li-Ion Battery, an I2S digital microphone and, of course, an Elsa-dress.
I had an ESP32 microcontroller laying around and used it as the core of the system: it supports I²S, has a floating point unit (FPU), is easy to use together with LED strips and has enough memory. The FPU makes it straightforward to use the same code on traditional computers as on embedded devices: fixed-point math can be avoided.
After soldering the components together and with the help from my better half to sew in the LED strip, it all came together. In the video below, the result of our work can be seen. The video first shows a song that should not and is not recognised. Then, “Let It Go” is played and correctly recognised. After the song is stopped, the lights go on for a while and finally stop: this is by design to allow gaps in recognition. Lastly, the song is continued and again correctly recognised.
The code went through several iterations and was expanded beyond the original scope and became a capable general purpose acoustic fingerprinting system with its many applications. Olaf performs quite well thanks to its resource friendly design and the use of PFFT and LMDB. Especially LMDB, a fast, B+-tree backed key value store with low storage overhead enables performant storage and lookups.
The GitHub does not contain an example for the ESP32. That code depends on the microcontroller, digital microphone and pins used and Olaf needs to be hacked to exhibit the requested behaviour. All in all that code is much less reusable (and sharable, testable, maintainable). I have, however, included a platformIO project for Olaf on ESP32 for reference.
WASM: Olaf in the browser
Olaf, being written in ANSI C, can run in the browser thanks to the Emscripten compiler. According to its website, Emscripten ‘…lets you run C and C++ on the web at near-native speed without plugins’ Combining the Web Audio API and the WASM version of Olaf makes web-based acoustic fingerprinting applications possible.
Below you can try out Olaf. The exact same code is running on your browser as on the ESP32 demonstrated above. This means that Olaf is listening to recognise ‘Let It Go’ from the Frozen soundtrack. For your convenience the song can be started below on the left. On the right, you can start Olaf by allowing incoming audio to be analysed. The FFT is calculated by Olaf and visualised using Pixi.js. After a few seconds the red fingerprints should become green, indicating a match. Once you stop the song, the fingerprints will eventually turn red again. As with the video above: going from a match to no match takes a couple of seconds to allow gaps in recognition.
1. Start the song and play it aloud. Singing along is encouraged.
2. Start the microphone and check whether recognition succeeds.
Olaf was featured on hackaday. There is also a small discussion about Olaf on Hacker News. A write-up of this project also ended up as a contribution to the Late Breaking Demo track of the first virtual ISMIR conference: Olaf ISMIR 2020 LBD abstract.
The audio shield takes care of the line level audio input. This audio input is then decoded. The decoding is done by libltc. The library runs as is on a Teensy without modification. The three elements are combined in a relatively simple teensy patch
To use the decoder connect the line level input left channel to an SMPTE source via e.g. an RCA plug.
During an experiment which monitors a music performance it might be a requirement to record music, video and sensor data synchronously. Recording analog sensors (balance boards, accelerometers, light sensors, distance sensors) together with audio and video is often problematic. Ideally standard DAW software can be used to record both audio and sensor data. A system is presented here that makes it relatively straightforward to record sensor data together with audio/video.
The basic idea is simple: a microcontroller is programmed to appear as a class compliant MIDI device. Analog measurements on the micro-controller are translated to a specific MIDI protocol. The MIDI data, on the capturing side, can then be converted again into the original sensor data. This setup has several advantages:
It makes it easy to record sensor data together with audio data in a standard DAW software package. Recording a recording a midi track and audio track simultaneously in, e.g., Ableton Live, is easy.
Communication with the micro-controller is bi-directional. The micro-controller can be programmed to react to certain MIDI messages. A note-on can, for example, be used to start analog sensor recording. These MIDI commands can be send from any possible source that can ‘speak’ MIDI.
Thanks to the Web MIDIAPI this construct presents an easy way to let analog sensors and websites interact.
Real time sonification of the sensor-data is also supported. There are many ready to go options to sonify MIDI. Axoloti, Max/MSP, Zupiter are some of the environments that can be pugged into.
Fig: Visualization in html of analog sensor data, captured as MIDI
While the concept is relatively simple, there are many details to get right. Please consult the MIDImorphosis github page which details the system that consists of an analog sensor, a MIDI protocol and a clocking infrastructure.
Together with Jeska, I presented an ongoing study on musical interaction. In the study one of the measurements was the body movement of two participants. This is done with boards that are equipped with weight sensors. The data that comes out of this can be inspected for synchronisation, quality and quantity of movement, movement periodicities.
The hardware is the work of Ivan Schepers, the software used to capture and transmit messages is called “the MIDImorphosis” and developed by me. The research is in collaboration with Jeska Buhman, Marc Leman and Alessandro Dell’Anna. An article with detailed findings is forthcoming.
The uniqueness of human music relative to speech and animal song has been extensively debated, but rarely directly measured. We applied an automated scale analysis algorithm to a sample of 86 recordings of human music, human speech, and bird songs from around the world. We found that human music throughout the world uniquely emphasized scales with small-integer frequency ratios, particularly a perfect 5th (3:2 ratio), while human speech and bird song showed no clear evidence of consistent scale-like tunings. We speculate that the uniquely human tendency toward scales with small-integer ratios may relate to the evolution of synchronized group performance among humans.
Automatic comparison of global children’s and adult songs
Music throughout the world varies greatly, yet some musical features like scale structure display striking crosscultural similarities. Are there musical laws or biological constraints that underlie this diversity? The “vocal mistuning” hypothesis proposes that cross-cultural regularities in musical scales arise from imprecision in vocal tuning, while the integer-ratio hypothesis proposes that they arise from perceptual principles based on psychoacoustic consonance. In order to test these hypotheses, we conducted automatic comparative analysis of 100 children’s and adult songs from throughout the world. We found that children’s songs tend to have narrower melodic range, fewer scale degrees, and less precise intonation than adult songs, consistent with motor limitations due to their earlier developmental stage. On the other hand, adult and children’s songs share some common tuning intervals at small-integer ratios, particularly the perfect 5th (~3:2 ratio). These results suggest that some widespread aspects of musical scales may be caused by motor constraints, but also suggest that perceptual preferences for simple integer ratios might contribute to cross-cultural regularities in scale structure. We propose a “sensorimotor hypothesis” to unify these competing theories.
At work we have a really nice piano and I wanted to be able to broadcast a live performance over the internet with low latency to potential live listeners. In all honesty, only my significant other gets moderately lukewarm about the idea of hearing me play live. Anyhow:
I did not find any practical tool to easily pump audio over the internet. I did find something that was very close called trx by Mark Hills: trx is a simple toolset for broadcasting live audio from Linux. It unfortunately only works with the ALSA audio system and is limited to Linux. I decided to extend it to support macOS and Pulse Audio. I also extended its name to form trix.
Audio Transmitter/Receiver over Ip eXchange (trix) is a simple toolset for broadcasting live audio from Linux or macOS. It sends and receives encoded audio over IP networks, via an audio interface. If audio interfaces are properly configured, a low-latency point-to-point or multicast broadband audio connection can be achieved. This could be used for networked music performances. The inclusion of the intermediate rtAudio library provides support for various audio input and outputs.
More information on trix can be found on the trix github page.
The system can be configured for low latency use. The whole chain is dependent several different components which each add to the total latency: audio input latency, encoder (algorithmic) delay, network latency and finally audio output latency.
Thanks to the use of RtAudio it should be possible to use low latency API’s to access audio devices (ASIO on windows or Jack on Unix). This means that audio input and output latencies can be as low as the hardware allows. The opus encoder/decoder that is used has a low algorithmic delay. By default it has a 25ms delay but it can be configured to only 2.5ms (see here). The network latency (and jitter) is very much dependent on the distance to cover. On a local network this can be kept low, when using wide area networks (the internet) control is lost and latencies can add up depending on the number of hops to take. Jitter can be problematic if the smallest possible buffers are used: then dropouts might occur and this might affect the audio in a noticeable way.
I have uploaded a small piece of software which allows users to find a specific audio marker in audio streams. It is mainly practical to synchronise a camera (audio/video) recording with other audio with the same marker. The marker is a set of three beeps. These three beeps are found with millisecond accurate precision within the audio streams under analysis. By comparing the timing of marker synchronization becomes possible. It can be regarded as an alternative for the movie clapper boards.
With the goal in mind to reduce common runner injuries we first need to measure some running style characteristics. Therefore, we have developed a sensor to measure how hard a runners foot repeatedly hits the ground. This sensor has been compared with laboratory equipment which proofs that its measurements are valid and can be repeated. The main advantages of our sensor is that it can be used ‘in the wild’, outside the lab on the runners regular tours. We want to use this sensor to provide real-time biofeedback in order to change running style and ultimately reduce injury risk.
Studies seeking to determine the effects of gait retraining through biofeedback on peak tibial acceleration (PTA) assume that this biometric trait is a valid measure of impact loading that is reliable both within and between sessions. However, reliability and validity data were lacking for axial and resultant PTAs along the speed range of over-ground endurance running. A wearable system was developed to continuously measure 3D tibial accelerations and to detect PTAs in real-time. Thirteen rearfoot runners ran at 2.55, 3.20 and 5.10 m*s-1 over an instrumented runway in two sessions with re-attachment of the system. Intraclass correlation coefficients (ICCs) were used to determine within-session reliability. Repeatability was evaluated by paired T-tests and ICCs. Concerning validity, axial and resultant PTAs were correlated to the peak vertical impact loading rate (LR) of the ground reaction force. Additionally, speed should affect impact loading magnitude. Hence, magnitudes were compared across speeds by RM-ANOVA. Within a session, ICCs were over 0.90 and reasonable for clinical measurements. Between sessions, the magnitudes remained statistically similar with ICCs ranging from 0.50 to 0.59 for axial PTA and from 0.53 to 0.81 for resultant PTA. Peak accelerations of the lower leg segment correlated to LR with larger coefficients for axial PTA (r range: 0.64–0.84) than for the resultant PTA per speed condition. The magnitude of each impact measure increased with speed. These data suggest that PTAs registered per stand-alone system can be useful during level, over-ground rearfoot running to evaluate impact loading in the time domain when force platforms are unavailable in studies with repeated measurements.