I have presented DiscStitch at the MIR get together at the Deezer headquarters in Paris.
DiscStitch is a solution to identify, align and mix digitized audio originating from (overlapping) laquer discs. The main contribution lays in the novel audio to audio alignment algorithm which is robust against some speed differences and variabilities.
I have updated Olaf – the Overly Lightweight Acoustic Fingerprinting system. Olaf is a piece of technology that uses digital signal processing to identify audio files by analyzing unique, robust, and compact audio characteristics – or “fingerprints”. The fingerprints are stored in a database for efficient comparison and matching. The database index allows for fast and accurate audio recognition, even in the presence of distortions, noise, and other variations.
Olaf is unique because it works on traditional computing devices, embedded microprocessors and in the browser. To this end tried to use ANSI C. C is a relatively small programming language but has very little safeguards and is full of exiting footguns. I enjoy the limitations of C: limitations foster creativity. I also made ample use of the many footguns C has to offer: buffer overflows, memory leaks, … However, with the current update I think most serious bugs have been found. Some of the changes to Olaf include:
Fixed a rather nasty array out of bounds bug. The bug remained elusive due to the fact that a segfault was rare on macOS. Linux seems to be more diligent in that regard.
Added a quick way to skip already indexed files. Which improves usability significantly when working with larger datasets.
Improved command line output and fixed incorrectly reported times. The reported start and stop time of a query was wrong and is now fixed.
Olaf now supports caching fingerprints in simple text files. This makes fingerprint extraction much faster since all cores of the system can be used to extract fingerprints and dump them to text files. Writing prints to the database from multiple threads is slow since they need to wait for access to the locked database. There is also a command to store all cashed fingerprints in a single go.
Added support for basic profiling with gprof. The profiler shows where optimizations can have the most impact.
Olaf now includes an algorithm for efficient max-filtering. The min-max filter algorithm by Daniel Lemire is implemented. The profiler showed that most time was spend during max-filtering: replacing the naive max-filter with the Lemire max-filter improved performance drastically.
CI with Github Actions which checks if checked in sources compile and tests some of the basic functionality automatically.
Updated the Zig build script for cross-compilation and updated the pre-build Windows version.
Tested the system with larger databases. The FMA-full datasets, which comprises almost a full year of audio was indexed and queried without problems on a single pc. The limits of Olaf with respect to indexed size is probably a few times larger.
Tested, fixed and improved the ‘memory database’ version. Also added documentation to the readme.
Made a basic web example to call the WASM version of Olaf.
Added an ESP32 example, showing how Olaf can run on this microprocessor. It runs without an external microphone but uses a test audio file. Previously some small changes were needed to Olaf to run on the ESP32, now the exact same code is used.
Anyhow, what originally started as a rather quick and dirty hack has been improved quite a bit. The takeaway message: in the world of software it does seem possible to polish a turd.
As a way to get to know the Rust programming language I have developed a couple of practical tools for OSC and MIDI debugging. OSC and MIDI are protocols which are almost always used for applications dealing with music. In these applications latency should be kept in check. Languages with garbage collection (Java, Go) and scripting languages (Ruby, Python, …) are hard to tune for low-latency applications and do not really have real-time guarantees. Rust, as a modern alternative for C/C++, is a better fit for cross platform CLI low-latency applications.
This opens a couple of possibilities which are discussed below.
Sending UDP messages from the browser
One of the ways to send OSC messages from a browser to a local network is by using the MIDI out capability of browsers and – using mot – translating MIDI to OSC an example can be seen below.
Fig: sending an UDP message to a network from a Browser using a the mot MIDI to OSC bridge, click the image for a better readable version.
Measuring UDP message latency
Both MIDI and OSC can be seen as rather general data encapsulation protocols with wide support in terms of libraries and cross platform support. Their value goes beyond mere musical applications. The same holds for mot. In this example we are using mot to measure UDP message latency between two hosts.
On the first host we send MIDI messages from MIDI device 0 over OSC to another host with e.g. mot midi_to_osc 192.168.1.12:3000 /m 0. At the other host we receive the OSC messages and send them to a virtual device: mot osc_to_midi 192.168.1.12:3000 /m 6666.
At the second host we return messages from the virtual device to the first host: mot midi_to_osc 192.168.1.4:5000 /m 1. Perhaps you first need to do mot midi_to_osc -l to find the index of the virtual device. As a final step the messages can be received at the first host and returned to the original midi device. On the first host: mot osc_to_midi 192.168.1.4:5000 /m 0.
If the original MIDI device is a Teensy running the “roundtrip patch” then finally the roundtrip time is accurately measured and shown in the serial console. I am sure the previous text is cromulent, totally not contrived and not confusing. Anyway, to make it more confusing: this is what happens when you use a single host to do midi to osc to midi to osc to midi and use the loopback networking device:
Fig: MIDI to OSC to MIDI to OSC to MIDI roundtrip latency.
Visualizing sensor data in the browser
Fig: Sensor data as MIDI.
When capturing sensor data on microcontrollers, data can be encoded into MIDI. This makes almost any sensor practically useful in Ableton Live or similar environments. It also makes it compatible with all other MIDI supporting devices. With mot it becomes trivial to send MIDI encoded sensor data over OSC e.g. to a central place to log that data.
Another use case is to visualize the incoming data in real-time. A single web page which reads and visualizes incoming MIDI-sensor data becomes much more useful if streams from other devices can be visualized as well with the mot midi_to_osc and mot osc_to_midi commands.
With the Rust compiler it is relatively easy to cross-compile for different targets. There is however an important limitation in mot. Windows has no support for virtual MIDI ports which limits the usefulness of mot on that platform.
Fig: Advances in Speech and Music Technology book cover.
I have recently published an chapter in an academic book published by Springer. The topic of the book is of interest to me but can be perceived as rather dry: Advances in Speech and Music Technology.
The chapter I co-authored presented two case studies on detecting duplicates in music archives. The fist case study deals with segmentation reuse in an archive of early electronic music. The second with meta-data reuse in an archive of a public broadcaster containing digitized commercial shellac disc recordings with many duplicates.
Duplicate detection being the main topic, I decided to title the article Duplicate Detection for for Digital Audio Archive Management. It is easy to miss, and not much is lost if you do, but there is a duplicate ‘for’ in the title. If you did detect the duplicate you have detected the duplicate in the duplicate detection article. Since I have fathered two kids I see it as an hard earned right to make dad-jokes like that. Even in academic writing.
It was surprisingly difficult to get the title published as-is. At every step of the academic publishing process (review, editorial, typesetting, lay-outing) I was asked about it and had to send an email like the one below. Every email and every explanation made my second-guess my sense of humor but I do stand by it.
To: Editors ASMT
I have updated my submission on easychair in…
I would like to keep the title however as is an attempt at word-play. These things tend to have less impact when explained but the article is about duplicate detection and is titled ‘Duplicate detection for for digital audio archive management’. The reviewer, attentively, detected the duplicate ‘for’ but unfortunately failed to see my attempt at humor. To me, it is a rather harmless witticism.
Anyway, I do think that humor can serve as a gateway to direct attention to rather dry, academic material. Also the message and the form of the message should not be confused. John Oliver, for example, made his whole career on delivering serious sometimes dry messages with heaps of humor: which does not make the topics less serious. I think there are a couple of things to be learned there. Anyway, now that I have your attention, please do read the author version of Duplicate Detection for for Digital Audio Archive Management: Two Case Studies.
TarsosDSP is a Java library for audio processing I have started working on more than 10 years ago. The aim of TarsosDSP is to provide an easy-to-use interface to practical music processing algorithms. Obviously, I have been using it myself over the years as my go-to library for audio-processing in Java. However, a number of gradual changes in the java ecosystem made TarsosDSP more and more difficult to use.
Since I have apparently not been the only one using it, there was a need to give it some attention. During the last couple of weeks I have found the time to give it this much needed attention. This resulted in a number of updates, some of the changes include:
Change of the build system from Apache Ant to Gradle
Make use of Java Modules to make TarsosDSP compatible with the ModulePath introduced in Java 9.
Packaged the software into a maven compatible format, which makes it easy to use as a dependency.
CI with GitHub actions to automatically build and test the software.
Updated some examples shipped with the TarsosDSP. I have still still some examples to verify.
Improved handling of errors on reading audio via ffmpeg
Fig: The updated TarsosDSP release contains many CLI and GUI example applications.
Notably the code of TarsosDSP has not changed much apart from some cosmetic changes. This backwards compatibility is one of the strong points of Java. With this update I am quite confident that TarsosDSP will also be usable during the next decade as well.
This blog has been running on Caddy for the last couple of months. Caddy is a http server with support for reverse proxies and automatic https. The automatic https feature takes care of requesting, installing and updating SSL certificates which means that you need much less configuration settings or maintenance compared with e.g. lighttpd or Nginx. The underlying certmagicACME client is responsible for requesting these certificates.
Before, it was using lighttpd but the during the last decade the development of lighttpd has stalled. lighttpd version 2 has been in development for 7 years and the bump from 1.4 to 1.5 has been taking even longer. lighttpd started showing its age with limited or no support for modern features like websockets, http/3 and finicky configuration for e.g. https with virtual domains.
Caddy with Ruby on Rails
I really like Caddy’s sensible defaults and the limited lines of configuration needed to get things working. Below you can find e.g. a reusable https enabled configuration for a Ruby on Rails application. This configuration does file caching, compression, http to https redirection and load balancing for two local application servers. It also serves static files directly and only passes non-file requests to the application servers.
If you are self-hosting I think Caddy is a great match in all but the most exotic or demanding setups. I definitely am kicking myself for not checking out caddy sooner: it could have saved me countless hours installing and maintaining https certs or configuring lighttpd in general.
JNI is a way to use C or C++ code from Java and allows developers to reuse and integrate C/C++ in Java software. In contrast to the Java code, C/C++ code is platform dependent and needs to be compiled for each platform/architecture. Also it is generally not a good idea to make users compile a C/C++ library: it is best provide precompiled libraries. As a developer it is, however, a pain to provide binaries for each platform.
With the dominance of x86 processors receding the problem of having to compile software for many platforms is becoming more pressing. It is not unthinkable to want to support, for example, intel and M1 macOS, ARM and x86_64 Linux and Windows. To support these platforms you would either need access to such a machine with a compiler or configure a cross-compiler for each system: both are unpractical. Typically setting up a cross-compiler can be time consuming and finicky and virtual machines can be tough to setup. There is however an alternative.
Zig is a programming language but, thanks to its support for C/C++, it also ships with an easy-to-use cross-compiler which is of interest here even if you have no intention to write a single line of Zig code. The built-in cross-compiler allows to target many platforms easily.
The Zig cross-compiler in practice
Cross compilation of C code is possible by simply replacing the gcc command with zig cc and adding a target argument, e.g. for targeting a Windows. There is more general information on zig as a cross-compiler here.
Cross-compiling a JNI library is not different to compiling other libraries. To make things concrete we will cross-compile a library from a typical JNI project: JGaborator packs the C/C++ library gaborator. In this case the C/C++ code does a computationally intensive spectral transformation of time domain data. The commands below create an x86_64 Windows DLL from a macOS with zig installed:
git clone --depth 1 https://github.com/JorenSix/JGaborator
zig cc -target x86_64-windows-gnu -c -O3 -ffast-math -fPIC pffft/pffft.c -o pffft/pffft.o
zig cc -target x86_64-windows-gnu -c -O3 -ffast-math -fPIC -DFFTPACK_DOUBLE_PRECISION pffft/fftpack.c -o pffft/fftpack.o
zig c++ -target x86_64-windows-gnu -I"pffft" -I"gaborator-1.7" $JNI_INCLUDES -O3\
-ffast-math -DGABORATOR_USE_PFFFT -o jgaborator.dll jgaborator.cc pffft/pffft.o pffft/fftpack.o
# jgaborator.dll: PE32+ executable (console) x86-64, for MS Windows
Note that, when cross-compiling from macOS, to target Windows a Windows JDK is needed. The windows JDK has other header files like jni.h. Some commands to download and use the JDK are commented out in the example above. Also note that targeting Linux from macOS seems to work with the standard macOS JDK. This is probably due to shared conventions regarding compilation of libraries.
To target other platforms, e.g. ARM Linux, there are only two things that need to be changed: the -target switch should be changed to aarch64-linux-gnu and the name of the output library should be (by Linux convention) changed to libjgaborator.so. During the build step of JGaborator a list of target platforms it iterated and a total of 9 builds are packaged into a single Jar file. There is also a bit of supporting code to load the correct version of the library.
Using a GitHub action or similar CI tools this cross compilation with zig can be automated to run on a software release. For Github the Setup Zig action is practical.
Loading the correct library
In a first attempt I tried to detect the operating system and architecture of the environment to then load the library but eventually decided against this approach. Mainly because you then need to keep an exhaustive list of supporting platforms and this is difficult, error prone and decidedly not future-proof.
In my second attempt I simply try to load each precompiled library limited to the sensible ones – only dll’s on windows – until a matching one is loaded. The rationale here is that the system itself knows best which library works and failing to load a library is computationally cheap. There is some code to iterate all precompiled libraries in a JAR-file so supporting an additional platform amounts to adding a precompiled library in the JAR folder: there is no need to be explicit in the Java code about architectures or OSes.
Trying multiple libraries has an additional advantage: this allows to ship multiple versions targeting the same architecture: e.g. one with additional acceleration libraries enabled and one without. By sorting the libraries alphabetically the first, then, should be the one with acceleration and the fallback without. In the case of JGaborator for mac aarch64 there is one compiled with -framework Accelerate and one compiled by the Zig cross-compiler without.
If you find yourself cross-compiling C or C++ for many platforms, consider the Zig cross-compiler. Even when you have no intention to write a single line of Zig code.
For JNI and Java the JGaborator source code might offer some inspiration to pre-compile and load libraries for many platforms with little effort.
CI tools can help to verify builds and automate Zig cross-compilation.
If you build for Windows make sure to include windows header-files even when there are no compilation errors using UNIX-header files.
Fig: Screenshot of Emotopa: a browser based tool to extract pitch organization from audio.
A couple of days ago I participated in the Music Hack Day – India. The event was organized the 10th and 11th of December in Bangaluru, India. During the event a representative of Smule suggested a task to evaluate the performance of karaoke-singers in terms of intonation. The idea was to employ pitch histogram like features to estimate pitch use of singers.
I offered to build a browser based application to extract pitch histograms from audio. At the end of the hack day I presented the first release of Emotopa with some limited functionality:
Next, a pitch detector runs on the audio and returns a list of pitch estimates.
Finally a histogram (technically a kernel density estimate) is constructed using the pitch estimates.
The user can export the pitch histogram, the pitch class histogram and the pitch annotations. These features successfully show the intonation quality of singers but the applications are much broader. Some potential applications have been described in (amongst others) the Tarsos article.
This year the ISMIR 2022 conference is organized from 4 to 9 December 2022 in Bengaluru, India. ISMIR is the main music technology and music information retrieval (MIR) conference. It is a relief to experience a conference in physical form and not through a screen.
I have contributed to the following work which is in the main paper track of ISMIR 2022:
Abstract: Audio Fingerprinting (AFP) is a well-studied problem in music information retrieval for various use-cases e.g. content-based copy detection, DJ-set monitoring, and music excerpt identification. However, AFP for continuous broadcast monitoring (e.g. for TV & Radio), where music is often in the background, has not received much attention despite its importance to the music industry. In this paper (1) we present BAF, the first public dataset for music monitoring in broadcast. It contains 74 hours of production music from Epidemic Sound and 57 hours of TV audio recordings. Furthermore, BAF provides cross-annotations with exact matching timestamps between Epidemic tracks and TV recordings. Approximately, 80% of the total annotated time is background music. (2) We benchmark BAF with public state-of-the-art AFP systems, together with our proposed baseline PeakFP: a simple, non-scalable AFP algorithm based on spectral peak matching. In this benchmark, none of the algorithms obtain a F1-score above 47%, pointing out that further research is needed to reach the AFP performance levels in other studied use cases. The dataset, baseline, and benchmark framework are open and available for research.
I have also presented a first version of DiscStitch, an audio-to-audio alignment algorithm. This contribution is in the ISMIR 2022 late breaking demo session:
Abstract: Before magnetic tape recording was common, acetate discs were the main audio storage medium for radio broadcasters. Acetate discs only had a capacity to record about ten minutes. Longer material was recorded on overlapping discs using (at least) two recorders. Unfortunately, the recorders used were not reliable in terms of recording speed, resulting in audio of variable speed. To make digitized audio originating from acetate discs fit for reuse, (1) overlapping parts need to be identified, (2) a precise alignment needs to be found and (3) a mixing point suggested. All three steps are challenging due to the audio speed variabilities. This paper introduces the ideas behind DiscStitch: which aims to reassemble audio from overlapping parts, even if variable speed is present. The main contribution is a fast and precise audio alignment strategy based on spectral peaks. The method is evaluated on a synthetic data set.
For example, Ghent University’s biblio and for the FWO’s academic profile do not allow to enter software as research output. The focus is still solely on papers, even when custom developed research software has become a fundamental aspect in many research areas. My role is somewhere between that of a ‘pure’ researcher and that of a research software engineer which makes this focus on papers quite relevant to me.
The paper aims to make the recent development on Panako‘count’. Thanks to the JOSS review process the Panako software was improved considerably: CI, unit tests, documentation, containerization,… The paper was a good reason to improve on all these areas which are all too easy to neglect. The paper itself is a short, rather general overview of Panako:
“Panako solves the problem of finding short audio fragments in large digital audio archives. The content based audio search algorithm implemented in Panako is able to identify a short audio query in a large database of thousands of hours of audio using an acoustic fingerprinting technique.”
I have been lucky to have been involved in an interdisciplinary research project around the low impact runner: a music based bio-feedback system to reduce tibial shock in over-ground running. In the beginning of October 2022 the PhD defence of Rud Derie takes place so it is a good moment to look back to this collaboration between several branches of Ghent University: IPEM , movement and sports science and IDLab.
The idea behind the project was to first select runners with a high foot-fall impact. Then an intervention would slightly nudge these runner to a running style with lower impact. A lower repetitive impact is expected to reduce the chance on injuries common for runners. A system was invented in which musical bio-feedback was given on the measured impact. The schema to the right shows the concept.
I was involved in development of the first hardware prototypes which measured acceleration on the legs of the runner and the development of software to receive and handle these measurement on a tablet strapped to a backpack the runner was wearing. This software also logged measurements, had real-time visualisation capabilities and allowed remote control and monitoring over the network. Finally measurements were send to a Max/MSP sonification engine. These prototypes of software and hardware were replaced during a valorization project but some parts of the software ended up in the final Android application.
Video: the left screen shows the indoor positioning system via UWB (ultra-wide-band) and the right screen shows the music feedback system and the real time monitoring of impact of the runner. Video by Pieter Van den Berghe
Over time the first wired sensors were replaced with wireless Bluetooth versions. This made the sensors easy to use and also to visualize sensor values in the browser thanks to the Web Bluetooth API. I have experimented with this and made two demos: a low impact runner visualizer and one with the conceptual schema.
Vid: Visualizing the Bluetooth Low Impact Runner sensor in the browser.
The following three studies shows a part of the trajectory of the project. The first paper is a validation of the measurement system. Secondly a proof-of-concept study is done which finally greenlights a larger scale intervention study.
Van den Berghe, P., Six, J., Gerlo, J., Leman, M., & De Clercq, D. (2019). Validity and reliability of peak tibial accelerations as real-time measure of impact loading during over-ground rearfoot running at different speeds. Journal of Biomechanics, 86, 238-242.
Van den Berghe, P., Lorenzoni, V., Derie, R., Six, J., Gerlo, J., Leman, M., & De Clercq, D. (2021). Music-based biofeedback to reduce tibial shock in over-ground running: A proof-of-concept study. Scientific reports, 11(1), 1-12.
Van den Berghe, P., Derie, R., Bauwens, P., Gerlo, J., Segers, V., Leman, M., & De Clercq, D. (2022). Reducing the peak tibial acceleration of running by music‐based biofeedback: A quasi‐randomized controlled trial. Scandinavian Journal of Medicine & Science in Sports
There are quite a number of other papers but I was less involved in those. The project also resulted in two PhD’s:
Motor retraining by real-time sonic feedback: understanding strategies of low impact running (2021) by Pieter Van den Berghe
Running on good vibes: music induced running-style adaptations for lower impact running (2022) by Rud Derie
I have created a web application to LTC.wasm decodes SMPTE timecodes from an LTC encoded audio signal.
To synchronize multiple music and video recordings a shared SMPTE timecode signal is often used. For practical purposes the timecode signal is encoded in an audio stream. The timecode can then be recorded in sync with microphone inputs or added to a video recording. The timecode is encoded in audio with LTC, linear timecode. A special decoder is needed to extract SMPTE timecode from the audio. This is exactly what the LTC.wasm application does.
The advantage of the web-based version versus the command line ltc-tools is that it does not need to be installed separately and that ffmpeg decodes audio. This means that almost any multimedia format is supported automatically. The command line version only supports a limited number of audio formats.
From an incoming media-file audio is extracted, downmixed to mono and and resampled. This is done with ffmpeg.audio.wasm a wasm version of ffmpeg.
For each audio track, fingerprints are extracted. These fingerprints reduce the the search space for alignment drastically.
Each list of fingerprints is aligned with the list of fingerprints from the reference. Resulting in a rough alignment
Cross correlation is done to refine the alignment resulting in sample accurate results.
Fig: media synchronization with audio-to-audio alignment.
It supports small time-scale adjustments of around 5%: audio alignment can still be found if audio speed differs a bit.
Some potential use cases where it might be of use:
To stitch partially overlapping audio recordings together resulting in a single long audio recording.
To synchronize multiple independent video recordings of the same event each with an audio recording of the environment.
To align a high quality microphone recording with video/low-quality audio recording of the same event. The low quality audio recorded with a camera can then be replaced with the high quality microphone audio.
Fig: stable diffusion imagining a networked music performance
This post describes how to send audio over a network using the ffmpeg suite. Ffmpeg is the Swiss army knife for working with audio and video formats. It is a command line tool that supports almost all audio formats known to man and woman. ffmpeg also supports streaming media over networks.
Here, we want to send audio recorded by a microphone, over a network to a single receiver on the other end. We are not aiming for low latency. Also the audio is going only in a single direction. This can be of interest for, for example, a networked music performance. Note that ffmpeg needs to be installed on your system.
The receiver – Alice
For the receiver we use ffplay, which is part of the ffmpeg tools. The command instructs the receiver to listen to TCP connections on a randomly chosen port 12345. The \?listen is important since this keeps the program waiting for new connections. For streaming media over a network the stateless UDP protocol is often used. When UDP packets go missing they are simply dropped. If only a few packets are dropped this does not cause much harm for the audio quality. For TCP missing packets are resent which can cause delays and stuttering of audio. However, TCP is much more easy to tunnel and the stuttering can be compensated with a buffer. Using TCP it is also immediately clear if a connection can be made. With UDP packets are happily sent straight to the void and you need to resort to wiresniffing to know whether packets actually arrive.
In this example we use MPEGTS over a plain TCP socket connection. Alteratively RTMP could be used (which also works over TCP). RTP , however is usually delivered over UDP.
The shorthand address 0.0.0.0 is used to bind the port to all available interfaces. Make sure that you are listening to the correct interface if you change the IP address.
The sender – Björn
Björn, aka Bob, sends the audio. First we need to know from which microphone to use. To that end there is a way to list audio devices. In this example the macOS avfoundation system is used. For other operating systems there are similar provisions.
ffmpeg -f avfoundation -list_devices true -i ""
Once the index of the device is determined the command below sends incoming audio to the receiver (which should already be listening on the other end). The audio format used here is MP3 which can be safely encapsulated into mpegts.
Note that the IP address 192.168.x.x needs to be changed to the address of the receiver. Now if both devices are on the same network the incoming audio from Bob should arrive at the side of Alice.
If sender and receiver are not on the same network it might be needed to do Network Addres Translation (NAT) and port forwarding. Alternatively an ssh tunnel can be used to forward local tcp connections to a remote location. So on the sender the following command would send the incoming audio to a local port:
The connection to the receiver can be made using a local port forwarding tunnel. With ssh the TCP traffic on port 12345 is forwarded to the remote receiver via an intermediary (remote) host using the following command:
LMDB is a fast key value store, ideal to store and query sorted data with small keys and values. LMDB is a pure C library but often used from other programming languages via some type of bindings. These bindings are ‘bridges’ between languages and are automatically present on supported platform. On new or unsupported platforms, however, you need to build a this bridge yourself.
This blog post is about getting java-lmdb working on such unsupported platform: arm64. The arm64 platform is much more popular since the introduction of the Apple silicon – M1 platform. On Apple M1 the default architecture of Docker images is also aarch64.
Next you need to build the lmdb library for your platform and copy it to a location where Java looks for it. This only works when compilers are already available on your system. In macOS you might need to install the XCode command line tools:
git clone --depth 1 https://git.openldap.org/openldap/openldap.git
make -e SOEXT=.dylib
cp liblmdb.dylib ~/Library/Java/Extensions
On Debian aarch64 the procedure is similar but a different extension is used (.so):
#apt install build-essential
git clone --depth 1 https://git.openldap.org/openldap/openldap.git
mv liblmdb.so /lib
Finally, to use the library in a JAR-file is might be needed to allow lmdbjava to access some parts of the JRE:
I have been lucky to be part of a fruitful interdisciplinary scientific collaboration around AMPEL: ‘The Augmented Movement Platform For Embodied Learning’. The recent publication of an article is an ideal occasion to give a glimpse behind the scenes.
Fig: Schematic representation of AMPEL, a floor with interactive tiles.
Around 2016 the idea arose to search for new potential rehabilitation approaches for persons with multiple sclerosis. Multiple sclerosis causes problems, in varying degrees, with both motor and cognitive function. Common rehabilitation approaches either work on motor or cognitive function. The idea (by Lousin Moumdjian, Marc Leman, Peter Feys) was to combine both motor and cognitive rehabilitation in a single combined ‘embodied learning’ paradigm.
After some discussion we wanted to perform a combined short-term memory and walking task. First the participants would be presented with a target trajectory which would then be performed by walking. During walking we would modulate feedback types (melodic, sounds or visual). To this end, an ‘intelligent floor’ device was needed that was able to present a target trajectory, register a performed trajectory and provide several types of feedback. After a search for off-the-shelf solutions it became clear that a custom hard-and-software platform was required.
After a great deal of cardboard prototyping we settled on a design consisting of interactive tiles. Thomas Vervust of UGhent NamiFab designed a PCB with force sensitive resistors (FSR) on the bottom and RGB LED’s on top. Ivan Schepers provided practical insights during prototyping and developed the hardware around the interactive tiles. I was responsible for programming the system. Custom software was developed for the tiles, a controller to drive the tiles and to run and record experiments. Finally the system was moved to a hospital where the experiments took place. To know more about the exact experiments, please read the following three publications on AMPEL:
However, browsers only support a small subset of audio formats and container formats. Dealing with many (legacy) audio formats is often a rather painful experience since there are so many media container formats which can contain a surprising variation of audio (and video) encodings. In short, decoding audio for in-browser analysis or playback is often problematic.
Luckily there is FFmpeg which claims to be ‘a complete, cross-platform solution to record, convert and stream audio and video’. It is, indeed, capable to decode almost any audio encoding known to man from about any container. Additionally, it also contains tools to filter, manipulate, resample, stretch, … audio. FFmpeg is a must-have when working with audio. It would be ideal to have FFmpeg running in a browser…
Next to the pure functionality of ffmpeg there are general advantages to run audio analysis software in the browser at client-side:
Ease-of-use: no software needs to be installed. The runtime comes with a compatible browser.
Privacy: Since media files are not transferred it is impossible for the system running the service to make unauthorised copies of these files. There is no need to trust the service since all processing happens locally, in the browser.
Speed: Downloading and especially uploading large media files takes a while. When files are kept locally, processing can start immediately and no time is wasted sending bytes over the internet. This results in a snappy user experience.
Computational load: the computational load of transcoding is distributed over the clients and not centralised on a (single) server. The server does not do any computing and only serves static files, so it can handle as many concurrent clients as its bandwidth allows.
PFFFT is a small, pretty fast FFT library programmed in C with a BSD-like license. I have taken it upon myself to compile a WebAssembly version of PFFFT to make it available for browsers and node.js environments. It is called pffft.wasm and available on GitHub.
The pffft.wasm library comes in two flavours. One is compiled with SIMD instructions while the other comes without these instructions. SIMD stands for ‘single instruction, multiple data’ and does what it advertises: in a single step it processes multiple datapoints. The aim of SIMD is to make calculations several times faster. Especially for workloads where the same calculations are repeated over and over again on similar data, SIMD optimisation is relevant. FFT calculation is such a workload.
This work presents updates to Panako, an acoustic fingerprinting system that was introduced at ISMIR 2014. The notable feature of Panako is that it matches queries even after a speedup, time-stretch or pitch-shift. It is freely available and has no problems indexing and querying 100k sea shanties. The updates presented here improve query performance significantly and allow a wider range of time-stretch, pitch-shift and speed-up factors: e.g. the top 1 true positive rate for 20s query that were sped up by 10 percent increased from 18% to 83% from the 2014 version of Panako to the new version. The aim of this short write-up is to reintroduce Panako, evaluate the improvements and highlight two techniques with wider applicability. The first of the two techniques is the use of a constant-Q non-stationary Gabor transform: a fast, reversible, fine-grained spectral transform which can be used as a front-end for many MIR tasks. The second is how near-exact hashing is used in combination with a persistent B-Tree to allow some margin of error while maintaining reasonable query speeds.
Together with the paper there is also a poster and a short video presentation which explains the work:
Have you ever found yourself wondering how to build an accurate, low-latency LTC decoder with a common micro-controller? Well! Wonder no more and read on! Or, stop reading and do go read something that is more appealing to your predispositions.
SMPTE timecodes were originally used to synchronize audio and video material. SMPTE timecode data is often encoded into audio using LTC or linear time code. This special audio stream can be recorded together with other audio and video material. By decoding the LTC audio afterwards and working back to SMPTE timecodes, synchronization of multiple camera angles and audio material becomes straightforward. This concept tagging data streams with SMPTE timecodes is also used for other types of data.
Fig: LTC is a ‘self-clocking’ protocol for which a period can be found automatically. Once the period is found, transitions within the period are counted. A period with a transition translates to a 1, a period without any transitions to a 0.
SMPTE timecodes supports up to 30 frames per second and this resolution might not be sufficient for some data streams. It helps if the frames could be split up and 60 or 120 frames per second could be generated. With a low latency LTC decoder it would be possible to support this case and, for example, provide four pulses for every SMPTE frame. To be more precise: a SMPTE frame consists of 80 bits and in this case we would send a pulse exactly when decoding bit 0, bit 20, bit 40 and bit 60. We would then be able to sample at 120Hz while staying in sync with the SMPTE.
My first attempt was to treat the signal like audio and use a ready built library for LTC audio decoding The problem there is that sampling is done which might not exactly match the SMPTE bit transition period and relatively large buffers are used to decode LTC. The bit exact decoding is not possible using this method: the latency is too large, the method also uses excessive computational power and memory.
Fig: Biasing circuit to offset voltages
In my second attempt, the current iteration, interrupts are used to detect rising and falling edges in the LTC stream. By counting the number of microseconds between these edges a bit string is constructed. Effectively decoding LTC without any wasted computational power or memory and at a very low latency. If the LTC stream is well-formed, following each incoming bit and reacting to it becomes straightforward. Finally, after gently massaging the LTC bit string, SMPTE timecodes ooze out of the system at a low latency.
I have implemented a low latency LTC and SMPTE timecode data decoder for a Teensy microcontroller. One of the current limitations is that only 30fps SMPTE without skipped frames is supported. Another limitation is that the precision of the derived 120Hz clock is dependent on the sampling rate of the encoded audio signal: if e.g. only 8000Hz is used, transitions can only be precise up until 125µs. The derived clock will jitter slightly but will not drift.
There is still a slight problem with audio and Teensy input: audio is generally transmitted from -1.8V to +1.8V and not – as a Teensy would expect – from 0 to 3.3V. To make this change a small biasing circuit is placed before the Teensy input. In my case two 100k resistors and a 0.1uF capacitor worked best. The interrupt is relatively robust against signals that are a clipping (outside the 0 – 3.3V) or slightly too silent. If the signal becomes too small LTC decoding obviously fails.