Hi, I'm Joren. Welcome to my website. I'm a research software engineer in the field of Music Informatics and Digital Humanities. Here you can find a record of my research and projects I have been working on. Learn more »
For the opening of the Sport Science Laboratory – Jacques Rogge of University Ghent I have created a demo of a system to visualize running impact. The demo can be seen starting at 45s in the video below.
UGent
Collega Edith van Dyck stuurde vorige week een persbericht rond over het onderzoek dat ze deed rond muziek en sporten. UGent persbericht ‘Muziek beïnvloedt pasfrequentie bij lopers’:
Aangezien heel wat joggers met muziek trainen, wilden onderzoekers van het IPEM (het onderzoekscentrum van de afdeling Musicologie, Vakgroep Kunst-, Muziek-, en Theaterwetenschappen aan de UGent) nagaan of het tempo van muziek de pasfrequentie tijdens het lopen kan beïnvloeden. Eerdere studies hadden al aangetoond dat muziek een motiverend effect kan hebben op sportprestaties en dat een hogere pasfrequentie blessurepreventief kan werken.
Een neerslag van het onderzoek is te lezen in het artikel Spontaneous Entrainment of Running Cadence to Music Tempo. Het persbericht werd goed opgepikt door de media en ook de lokale televisiezender AVS vertoonde interesse. Een cameraploeg kwam langs en dit resulteerde in volgend verslag. In het verslag spelen mijn vriendin en ikzelf een figurantenrol. De hoofdrol is weggelegd voor Dieter.
UGent and Dutch
The Mi Band is a bracelet with some sensors, three RGB leds and a vibration motor. It is marketed as an activity tracker and notifier. It is a neat little device that communicates via Bluetooth LE and has a battery life of around 30 days. It would be nice if it could be used for whatever purpose you want but alas, its API is not very open. This blog post gives pointers to useful resources and tips to make it work with your own code.
There have been some efforts to reverse engineer the Bluetooth protocol. This blog post contains some info. There are even complete implementations available of the protocol, there is a Mi Band protocol implementation in python and a Mi Band protocol implementation in Java. It is however not always clear which firmware version is targeted.
I would advise against installing the official Mi Band app, if you want to use it with custom code. The app upgrades the firmware to the latest version and it seems that Xiaomi is obfuscating the protocol more and more with each version. I was able to send vibrate and led commands to a Mi Band with firmware version 10.0.9.3. With the previously mentioned sources and the flow described to the right the device reacts to commands. I used an Android device. The flow:
Some notes: the self-test command works without the set user step. For Android the Mi Band protocol implementation in Java works well. To check the firmware version of the device, call the get device info characteristic. The last bytes, interpreted as an integer, define the version info. For my device it is 10.9.3.2:
Write to characteristic 0000ff05-0000-1000-8000-00805f9b34fb onCharacteristicWrite status: 0 characteristic 0000ff05-0000-1000-8000-00805f9b34fb Read firmware version 11 value: 2 12 value: 3 13 value: 9 14 value: 0 15 value: 1
Another note: the set user info needs to be called with a 1 as type the first time the band is used. This is done with new UserInfo(20111111, 1, 32, 180, 55, "NM", 1) with the Android sdk by GitHub user pangliang. This sets and overwrites the user info. The next times you do not want to overwrite the info and the type needs to be zero.
UGent and Code
vibrate_flowchart__1_.svg
The article titled “Synchronizing Multimodal Recordings Using Audio-To-Audio Alignment” by Joren Six and Marc Leman has been accepted for publication in the Journal on Multimodal User Interfaces. The article will be published later this year. It describes and tests a method to synchronize data-streams. Below you can find the abstract, pointers to the software under discussion and an author version of the article itself.
Abstract: Research on the interaction between movement and music often involves analysis of multi-track audio, video streams and sensor data. To facilitate such research a framework is presented here that allows synchronization of multimodal data. A low cost approach is proposed to synchronize streams by embedding ambient audio into each data-stream. This effectively reduces the synchronization problem to audio-to-audio alignment. As a part of the framework a robust, computationally efficient audio-to-audio alignment algorithm is presented for reliable synchronization of embedded audio streams of varying quality. The algorithm uses audio fingerprinting techniques to measure offsets. It also identifies drift and dropped samples, which makes it possible to find a synchronization solution under such circumstances as well. The framework is evaluated with synthetic signals and a case study, showing millisecond accurate synchronization.
To read the article, consult the author version of Synchronizing Multimodal Recordings Using Audio-To-Audio Alignment. The data-set used in the case study is available here. It contains a recording of balanceboard data, accelerometers, and two webcams that needs to be synchronized. The final publication is available at Springer via 10.1007/s12193-015-0196-1
The algorithm under discussion is included in Panako an audio fingerprinting system but is also available for download here. The SyncSink application has been packaged separately for ease of use.
To use the application start it with double click the downloaded SyncSink JAR-file. Subsequently add various audio or video files using drag and drop. If the same audio is found in the various media files a time-box plot appears, as in the screenshot below. To add corresponding data-files click one of the boxes on the timeline and choose a data file that is synchronized with the audio. The data-file should be a CSV-file. The separator should be ‘,’ and the first column should contain a time-stamp in fractional seconds. After pressing Sync a new CSV-file is created with the first column containing correctly shifted time stamps. If this is done for multiple files, a synchronized sensor-stream is created. Also, ffmpeg commands to synchronize the media files themselves are printed to the command line.
This work was supported by funding by a Methusalem grant from the Flemish Government, Belgium. Special thanks goes to Ivan Schepers for building the balance boards used in the case study. If you want to cite the article, use the following BiBTeX:
@article{six2015multimodal,
author = {Joren Six and Marc Leman},
title = {{Synchronizing Multimodal Recordings Using Audio-To-Audio Alignment}},
issn = {1783-7677},
volume = {9},
number = {3},
pages = {223-229},
doi = {10.1007/s12193-015-0196-1},
journal = {{Journal of Multimodal User Interfaces}},
publisher = {Springer Berlin Heidelberg},
year = 2015
}
Research papers
SyncSink-1.4.jar, 2015.synchronized-recording.pdf, and SyncDataset.zip
This post explains how to do real-time pitch-shifting and audio time-stretching in Java. It uses two components. The first component is a high quality software C++ library for audio time-stretching and pitch-shifting C++ called Rubber Band. The second component is a Java audio library called TarsosDSP. To bridge the gap between the two JNI is used. Rubber Band provides a JNI interface and starting from the currently unreleased version 1.8.2, makefiles are provided that make compiling and subsequently using the JNI version of Rubber Band relatively straightforward.
However, it still requires some effort to control real-time pitch-shifting and audio time-stretching from java. To make it more easy some example code and documentation is available in a GitHub repository called RubberBandJNI. It documents some of the configuration steps needed to get things working. It also offers precompiled libraries and documents how to compile those for the following systems:
If the instructions are followed rather precisely you are able to control the tempo of a song in real-time with the following Java code:
1
2
3
4
5
6
7
8
float tempoFactor = 0.8f;
float pitchFactor = 1.0f;
AudioDispatcher adp = AudioDispatcherFactory.fromPipe("music.mp3", 44100, 4096, 0);
TarsosDSPAudioFormat format = adp.getFormat();
rbs = new RubberBandAudioProcessor(44100, tempoFactor, pitchFactor);
adp.addAudioProcessor(rbs);
adp.addAudioProcessor(new AudioPlayer(JVMAudioInputStream.toAudioFormat(format)));
new Thread(adp).start();
Code, TarsosDSP, and UGent
This post describes how to decode MP3’s using an already compiled ffmpeg binary on android. Using ffmpeg to decode audio on Android has advantages:
mkv, avi, mov,… just works.The main disadvantage is that you need an ffmpeg build for your Android device. Luckily some poor soul already managed to compile ffmeg for Android for several architectures. The precompiled ffmpeg binaries for Android are available for download and are mirrored here as well.
To bridge the ffmpeg binary and the java world TarsosDSP contains some glue code. The AndroidFFMPEGLocator is responsible to find and extract the correct binary for your Android device. It expects these ffmpeg binaries in the assets folder of your Android application. When the correct ffmpeg binary has been extracted and made executable the PipeDecoder is able to call it. The PipeDecoder calls ffmpeg so that decoded, downmixed and resampled PCM samples are streamed into the Java application via a pipe, which explains its name.
With the TarsosDSP Android library the following code plays an MP3 from external storage:
1
2
3
4
5
6
7
8
9
10
11
12
new AndroidFFMPEGLocator(this);
new Thread(new Runnable() {
@Override
public void run() {
File externalStorage = Environment.getExternalStorageDirectory();
File mp3 = new File(externalStorage.getAbsolutePath() , "/audio.mp3");
AudioDispatcher adp;
adp = AudioDispatcherFactory.fromPipe(mp3.getAbsolutePath(),44100,5000,2500);
adp.addAudioProcessor(new AndroidAudioPlayer(adp.getFormat(),5000, AudioManager.STREAM_MUSIC));
adp.run();
}
}).start();
This code just works if the application has the READ_EXTERNAL_STORAGE permission, includes a recent TarsosDSP-Android.jar, is ran on one of the supported ffmpeg architectures and has these binaries available in the assets folder.
UGent and TarsosDSP
This post describes a tool to quickly visualize and record analog signals with a Teensy micro-controller and some custom software. It is mainly useful to quickly get an idea of how an analog sensor reacts to different stimuli. Since it is also able to capture and store analog input siginals it is also useful to generate test data recordings which then can be used for example to test a peak detection algorithm on. The tool is called TeensyDAQ hinting at the Data AcQuisition features and the micro-controller used.
Some of the features of the TeensyDAQ:
The system consists of two parts. A hardware and a software part. The hardware is a Teensy micro-controller running an Arduino sketch that ready analog input A0 to A4 at the requested sampling rate. A Teensy is used instead of a regular Arduino for two reasons. First the Teensy is capable of much higher data throughput, it is able to send five reading at 8000Hz, which is impossible on Arduino. The second reason is the 13bit analog read resolution. Classic Arduino only provides 10 bits.
The software part reads data from the serial port the Teensy is attached to. It interprets the data and stores it in an efficient data-structure. As quickly as possible the data is visualized. The software is written in Java. A recent Java runtime environment is needed to execute it.
Try out the latest version of TeensyDAQ or check out the source code on the github TeensyDAQ source repository.
Code and UGent
TeensyAnalogRead.ino and TeensyDAQ-1.0.jar
TarsosDSP, the is a real-time audio processing library written in Java, is featured in EFY Plus Magazine of July 2015. It is a leading electronics magazine with a history going back more than 40 years and about 300 000 subscribers mainly in India. The index mentions this:
TarsosDSP: A Real-Time Audio Analysis and Processing Framework
In last month’s EFY Plus, we discussed Essentia, a C++ library for audio analysis. In this issue we will discuss a Java based real-time audio analysis and processing framework known as TarsosDSP
To read the full article, buy a (digital) copy of the magazine.
UGent and TarsosDSP
This post describes how to get notifications from a Bluetooth LE or Bluetooth v4.0 device on a Linux machine. Since it took me a while to get it going it is perhaps of interest to others.
The hardware I used is an RFduino board and a Belikin mini Bluethooth v4.0 adapter. The RFduino was programmed to wait for an event with RFduino_pinWake(pni, HIGH). When the pin is HIGH a count is incremented and this number is send to any device that is listening. In my case a Linux machine. The code is essentially the same as the button example included in the RDduino software distribution.
To install the Bluetooth stack on Debian the following command is executed sudo apt-get install bluetooth bluez bluez-utils bluez-firmware. A blog post describes more about the Bluetooth tools. Some other interesting reads are Get started with Bluetooth Low Energy and this stackoverflow question. Once the stack is installed correctly the lescan utility should give an output like this:
1
2
3
4
$ sudo hcitool lescan
LE Scan ...
DC:87:CC:18:14:A5 RFduino
DC:87:CC:18:14:A5 (unknown)
Bluetooth LE works with the Generic Attribute Profile (GATT). A Bluetooth LE device can provide services by combining characteristics. These characteristics are the way to communicate with the device. Some characteristics are writable and are able to send notifications. To receive notifications one such characteristic (referred to with a hex handle) needs to be written. Write 0100 to get notifications, 0200 for indications (indications are notifications that are acknowledged), 0300 for both, or 0000 for nothing (default). With this in mind, the following command enables listening for notifications:
1
gatttool --device=DC:87:CC:18:14:A5 --char-write-req --handle=0x000f --value=0300 --listen
With those commands working, the process can be automated with a Ruby script to get Bluetooth LE notifications. The script essentially calls gatttool with the correct parameters and parses and reacts to its output. To make it work lescan needs to be called before starting the script:
1
2
3
4
5
6
7
8
9
10
11
$ sudo hcitool lescan && ruby bluetooth_notifications.rb
LE Scan ...
DC:87:CC:18:14:A5 RFduino
DC:87:CC:18:14:A5 (unknown)
Characteristic value was written successfully
Notification handle = 0x000e value: 41 decimal value: 65
Notification handle = 0x000e value: 42 decimal value: 66
Notification handle = 0x000e value: 43 decimal value: 67
Notification handle = 0x000e value: 44 decimal value: 68
Notification handle = 0x000e value: 45 decimal value: 69
Notification handle = 0x000e value: 46 decimal value: 70
This post describes how to connect music in your library with precomputed features. Say, for example, you are developing a DJ application and you want to facilitate mixing tracks. To provide a seamless mix you perhaps want information about beats and about the key the music in your library is in. Since vast databases of features are already available you probably want to access those, instead of using your own feature extractors and database. The problems that need to be addressed are:
To help with these task there are several open source tools and services available.
To identify music a condensed representation of musical audio is created. This process is known as acoustic fingerprinting. On the website AcoustID a tool is available to create such fingerprint. The library is called Chromaprint and the command line client is called fpcalc. Currently the latest version is Chromaprint version 1.2 and static binaries for fpcalc are available on the AcoustID website. A packages for Debian (and probably Ubuntu) can be installed by calling apt-get install libchromaprint-tools. Once this tool is correctly installed a fingerprint for a piece of music can be created:
1
2
3
4
5
fpcalc music.mp3
FILE=music.mp3
DURATION=168
FINGERPRINT=AQADtEmi..hADAAOCGAQghZRgQByjAEAICSMWYME
A fingerprint by itself is not of much use. The AcoustID webservice translates a fingerprint into one or more MusicBrainz identifiers. One fingerprint can result in multiple identifiers because the same audio can be released on several albums. There is documentation for AcoustID webservice available. To use the webservice an API key is needed. Confusingly, the AcoustID service has two types of API keys. One for end-users and one for developers. The last type is needed to translate ID’s. To request a developer API key, log in on the AcoustID website and “add an application”, there you can find the correct API key. Substitute dev_api_key in the following URL. Also change the fingerprint and duration to match the information provided by the fpcalc application. The webservice should reply with a set of MusicBrainz identifiers:
http://api.acoustid.org/v2/lookup?client=dev_api_key&duration=x&fingerprint=ADORIF...LKJE6&meta=recordingids
AcousticBrainz provides features for a subset of music that has a MusicBrainz identifier. Currently about a million tracks are analyzed but more are added every day. The API for the webservice is straightforward:
GET http://acousticbrainz.org/96685213-a25c-4678-9a13-abd9ec81cf35/low-level GET http://acousticbrainz.org/96685213-a25c-4678-9a13-abd9ec81cf35/high-level
The low-level features include beat positions and chroma information. For the hypothetical DJ-application this is the information that would be used.
If you find the services useful please consider contributing to MusicBrainz, AcoustID and AcousticBrainz.
A small Ruby script to automatically fetch features for audio can be downloaded here. It needs Ruby and a RubyGems to parse JSON. On Debian this can be installed with apt-get install ruby and rubygems install json. Once these dependencies are installed the script can be ran as follows:
1
2
3
4
5
6
7
8
ruby mbid_lookup.rb example.mp3
Found 6 musicbrainz identifiers!
Not found in AcousticBrainz: 0afcd4a1-3709-499b-b76f-0d5491f839a5
Beat positions for 3d49fab8-fd08-42be-b0d2-9f1dc884d902: 0.522448956966,1.05650794506,1.57895684242,2.10140585899,2.61224484444,3.13469386101
Not found in AcousticBrainz: 448258f0-aa5a-4968-8efd-8c9348d5142e
Not found in AcousticBrainz: adcd7079-57d9-49bd-a36b-a20fa27b02b1
Beat positions for d1cd1321-0b66-4848-935e-f3afba6c7356: 0.441179126501,0.905578196049,1.369977355,1.83437633514,2.29877543449,2.76317453384
Not found in AcousticBrainz: e1f433be-af6b-4b5d-a969-4b53f014c395