Fellow of the Audio Engineering Society

The Audio Engineering Society’s Fellowship Award is given to ‘a member who had rendered conspicuous service or is recognized to have made a valuable contribution to the advancement in or dissemination of knowledge of audio engineering or in the promotion of its application in practice’.

Today at the 147th AES Convention, I was given the Fellowship Award for valuable contributions to, and for encouraging and guiding the next generation of researchers in, the development of audio and musical signal processing.

This is quite an honour, of which I’m very proud. And it puts me in some excellent company. A lot of greats have become Fellows of the AES (Manfred SchroederVesa Valimaki, Poppy Crum, Bob Moog, Richard Heyser, Leslie Ann Jones, Gunther Thiele and Richard Small…) which also means I have a lot to live up to.

And thanks to the AES,

Josh Reiss

Advertisements

Radical and rigorous research at the upcoming Audio Engineering Society Convention

aes-ny-19-logo-small

We previewed the 142nd, 143rd, 144th  and 145th Audio Engineering Society (AES) Conventions, which we also followed with wrap-up discussions. Then we took a break, but now we’re back to preview the 147th AES  convention, October 16 to 19 in New York. As before, the Audio Engineering research team here aim to be quite active at the convention.

We’ve gathered together some information about a lot of the research-oriented events that caught our eye as being unusual, exceptionally high quality, involved in, attending, or just worth mentioning. And this Convention will certainly live up to the hype.

Wednesday October 16th

When I first read the title of the paper ‘Evaluation of Multichannel Audio in Automobiles versus Mobile Phones‘, presented at 10:30, I thought it was a comparison of multichannel automotive audio versus the tinny, quiet mono or barely stereo from a phone. But its actually comparing results of a listening test for stereo vs multichannel in a car, with results of a listening test for stereo vs multichannel for the same audio, but from a phone and rendered over headphones. And the results look quite interesting.

Deep neural networks are all the rage. We’ve been using DNNs to profile a wide variety of audio effects. Scott Hawley will be presenting some impressive related work at 9:30, ‘Profiling Audio Compressors with Deep Neural Networks.’

We previously presented work on digital filters that closely match their analog equivalents. We pointed out that such filters can have cut-off frequencies beyond Nyquist, but did not explore that aspect. ‘Digital Parametric Filters Beyond Nyquist Frequency‘, at 10 am, investigates this idea in depth.

I like a bit of high quality mathematical theory, and that’s what you get in Tamara Smyth’s 11:30 paper ‘On the Similarity between Feedback/Loopback Amplitude and Frequency Modulation‘, which shows a rather surprising (at least at first glance) equivalence between two types of feedback modulation.

There’s an interesting paper at 2pm, ‘What’s Old Is New Again: Using a Physical Scale Model Echo Chamber as a Real-Time Reverberator‘, where reverb is simulated not with impulse response recordings, or classic algorithms, but using scaled models of echo chambers.

At 4 o’clock, ‘A Comparison of Test Methodologies to Personalize Headphone Sound Quality‘ promises to offer great insights not just for headphones, but into subjective evaluation of audio in general.

There’s so many deep learning papers, but the 3-4:30 poster ‘Modal Representations for Audio Deep Learning‘ stands out from the pack. Deep learning for audio most often works with raw spectrogram data. But this work proposes learning modal filterbank coefficients directly, and they find it gives strong results for classification and generative tasks. Also in that session, ‘Analysis of the Sound Emitted by Honey Bees in a Beehive‘ promises to be an interesting and unusual piece of work. We talked about their preliminary results in a previous entry, but now they’ve used some rigorous audio analysis to make deep and meaningful conclusions about bee behaviour.

Immerse yourself in the world of virtual and augmented reality audio technology today, with some amazing workshops, like Music Production in VR and AR, Interactive AR Audio Using Spark, Music Production in Immersive Formats, ISSP: Immersive Sound System Panning, and Real-time Mixing and Monitoring Best Practices for Virtual, Mixed, and Augmented Reality. See the Calendar for full details.

Thursday, October 17th

An Automated Approach to the Application of Reverberation‘, at 9:30, is the first of several papers from our team, and essentially does something to algorithmic reverb similar to what “Parameter Automation in a Dynamic Range Compressor” did for a dynamic range compressor.

Why do public address (PA) systems sound for large venues sound so terrible? They actually have regulations for speech intelligibility. But this is only measured in empty stadiums. At 11 am, ‘The Effects of Spectators on the Speech Intelligibility Performance of Sound Systems in Stadia and Other Large Venues‘ looks at the real world challenges when the venue is occupied.

Two highlights of the 9-10:30 poster session, ‘Analyzing Loudness Aspects of 4.2 Million Musical Albums in Search of an Optimal Loudness Target for Music Streaming‘ is interesting, not just for the results, applications and research questions, but also for the fact that involved 4.2 million albums. Wow! And there’s a lot more to audio engineering research than what one might think. How about using acoustic sensors to enhance autonomous driving systems, which is a core application of ‘Audio Data Augmentation for Road Objects Classification‘.

Audio forensics is a fascinating world, where audio engineering is often applied to unusually but crucially. One such situation is explored at 2:15 in ‘Forensic Comparison of Simultaneous Recordings of Gunshots at a Crime Scene‘, which involves looking at several high profile, real world examples.

Friday, October 18th

There are two papers looking at new interfaces for virtual reality and immersive audio mixing, ‘Physical Controllers vs. Hand-and-Gesture Tracking: Control Scheme Evaluation for VR Audio Mixing‘ at 10:30, and ‘Exploratory Research into the Suitability of Various 3D Input Devices for an Immersive Mixing Task‘ at 3:15.

At 9:15, J. T. Colonel from our group looks into the features that relate, or don’t relate, to preference for multitrack mixes in ‘Exploring Preference for Multitrack Mixes Using Statistical Analysis of MIR and Textual Features‘, with some interesting results that invalidate some previous research. But don’t let negative results discourage ambitious approaches to intelligent mixing systems, like Dave Moffat’s (also from here) ‘Machine Learning Multitrack Gain Mixing of Drums‘, which follows at 9:30.

Continuing this theme of mixing analysis and automation is the poster ‘A Case Study of Cultural Influences on Mixing Preference—Targeting Japanese Acoustic Major Students‘, shown from 3:30-5, which does a bit of meta-analysis by merging their data with that of other studies.

Just below, I mention the need for multitrack audio data sets. Closely related, and also much needed, is this work on ‘A Dataset of High-Quality Object-Based Productions‘, also in the 3:30-5 poster session.

Saturday, October 19th

We’re approaching a world where almost every surface can be a visual display. Imagine if every surface could be a loudspeaker too. Such is the potential of metamaterials, discussed in ‘Acoustic Metamaterial in Loudspeaker Systems Design‘ at 10:45.

Another session, 9 to 11:30 has lots of interesting presentations about music production best practices. At 9, Amandine Pras presents ‘Production Processes of Pop Music Arrangers in Bamako, Mali‘. I doubt there will be many people at the convention who’ve thought about how production is done there, but I’m sure there will be lots of fascinating insights. This is followed at 9:30 by ‘Towards a Pedagogy of Multitrack Audio Resources for Sound Recording Education‘. We’ve published a few papers on multitrack audio collections, sorely needed for researchers and educators, so its good to see more advances.

I always appreciate filling the gaps in my knowledge. And though I know a lot about sound enhancement, I’ve never dived into how its done and how effective it is in soundbars, now widely used in home entertainment. So I’m looking forward to the poster ‘A Qualitative Investigation of Soundbar Theory‘, shown 10:30-12. From the title and abstract though, this feels like it might work better as an oral presentation. Also in that session, the poster ‘Sound Design and Reproduction Techniques for Co-Located Narrative VR Experiences‘ deserves special mention, since it won the Convention’s Best Peer-Reviewed Paper Award, and promises to be an important contribution to the growing field of immersive audio.

Its wonderful to see research make it into ‘product’, and ‘Casualty Accessible and Enhanced (A&E) Audio: Trialling Object-Based Accessible TV Audio‘, presented at 3:45, is a great example. Here, new technology to enhance broadcast audio for those with hearing loss iwas trialed for a popular BBC drama, Casualty. This is of extra interest to me since one of the researchers here, Angeliki Mourgela, does related research, also in collaboration with BBC. And one of my neighbours is an actress who appears on that TV show.

I encourage the project students working with me to aim for publishable research. Jorge Zuniga’s ‘Realistic Procedural Sound Synthesis of Bird Song Using Particle Swarm Optimization‘, presented at 2:30, is a stellar example. He created a machine learning system that uses bird sound recordings to find settings for a procedural audio model. Its a great improvement over other methods, and opens up a whole field of machine learning applied to sound synthesis.

At 3 o’clock in the same session is another paper from our team, Angeliki Mourgela presenting ‘Perceptually Motivated Hearing Loss Simulation for Audio Mixing Reference‘. Roughly 1 in 6 people suffer from some form of hearing loss, yet amazingly, sound engineers don’t know what the content will sound like to them. Wouldn’t it be great if the engineer could quickly audition any content as it would sound to hearing impaired listeners? That’s the aim of this research.

About three years ago, I published a meta-analysis on perception of high resolution audio, which received considerable attention. But almost all prior studies dealt with music content, and there are good reasons to consider more controlled stimuli too (noise, tones, etc). The poster ‘Discrimination of High-Resolution Audio without Music‘ does just that. Similarly, perceptual aspects of dynamic range compression is an oft debated topic, for which we have performed listening tests, and this is rigorously investigated in ‘Just Noticeable Difference for Dynamic Range Compression via “Limiting” of a Stereophonic Mix‘. Both posters are in the 3-4:30 session.

The full program can be explored on the Convention Calendar or the Convention website. Come say hi to us if you’re there! Josh Reiss (author of this blog entry), J. T. Colonel, Angeliki Mourgela and Dave Moffat from the Audio Engineering research team within the Centre for Digital Music, will all be there.

Congratulations, Dr. Will Wilkinson

This afternoon one of our PhD student researchers, Will Wilkinson, successfully defended his PhD. The form of these exams, or vivas, varies from country to country, and even institution to institution, which we discussed previously. Here, its pretty gruelling; behind closed doors, with two expert examiners probing every aspect of the PhD.
Will’s PhD was on ‘Gaussian Process Modelling for Audio Signals.’

Audio signals are characterised and perceived based on how their spectral make-up changes with time. Latent force modelling assumes these characteristics come about as a result of a common input function passing through some input-output process. Uncovering the behaviour of these hidden spectral components is at the heart of many applications involving sound, but is an extremely difficult task given the infinite number of ways any signal can be decomposed.

Will’s thesis studies the application of Gaussian processes to audio, which offer a way to specify probabilities for these functions whilst encoding prior knowledge about sound, the way it behaves, and the way it is perceived. Will advanced the theory considerably, and tested his approach for applications in sound synthesis, denoising and source separation tasks, among others.

http://c4dm.eecs.qmul.ac.uk/audioengineering/latent-force-synthesis/ – demonstrates some of his research applied to sound synthesis, and https://fxive.com/app/main-panel/Mammals.html is a real-time demonstration of his Masters work on sound synthesis for mammalian vocalisations.

Here’s a list of all Will’s papers while a member of the Intelligent Sound Engineering team and the Machine Listening Lab.

Back end developer needed for sound synthesis start-up

logo black on white

FXive (fxive.com) is a real-time sound effect synthesis framework in the browser, spun-out from research developed at Queen Mary University of London by the team behind this blog. It is currently front-end only. We’d like to subcontract a backend developer to implement:

  • Sign-up, log-in and subscription system
  • Payment system for subscription, which offers unlimited sound downloads, and purchasing sounds individually

Additional functionalities can be discussed.

If you’re interested or know a web developer that might be interested, please get in touch with us from fxiveteam@gmail.com .

You can check out some of the sound effect synthesis models used in FXive in previous blog entries.

 

@c4dm @QMUL #backend #webdeveloper #nodejs

 

Nonlinear Audio Effects at ICASSP 2019

The Audio Engineering research team within the Centre for Digital Music is going to be present at ICASSP 2019.

Marco Martínez is presenting the paper ‘Modeling Nonlinear Audio Effects With End-to-end Deep Neural Networks‘, which can be found here.

Basically, given that nonlinear audio effects are widely used by musicians and sound engineers and taking into account that most existing methods for nonlinear modeling are often either simplified or optimized to a very specific circuit. In this work, we introduce a general-purpose deep learning architecture for generic black-box modeling of nonlinear and linear audio effects.

We show the model performing nonlinear modeling for distortion, overdrive, amplifier emulation and combination of linear and nonlinear audio effects.

jimi.jpg

You can listen to some audio samples here.

Details about the presentation:

Session: AASP-L6: Music Signal Analysis, Processing and Synthesis
Location: Meeting Room 1
Time: Thursday, May 16, 09:20 – 09:40 (Approximate)

Title: Modeling Nonlinear Audio Effects With End-to-end Deep Neural Networks
Authors: Marco A. Martinez Ramirez, Joshua D. Reiss

Cool sound design and audio effects projects

Every year, I teach two classes (modules), Sound Design and Digital Audio Effects. In both classes, the final assignment involves creating an original work that involves audio programming and using concepts taught in class. But the students also have a lot of free reign to experiment and explore their own ideas. Last year, I had a well received blog entry about the projects.

The results are always great. Lots of really cool ideas, many of which could lead to a publication, or would be great to listen to regardless of the fact that it was an assignment. Here’s a few of the projects this year.

From the Sound Design class;

  • A truly novel abstract sound synthesis (amplitude and frequency modulation) where parameters are controlled by pitch recognition and face recognition machine learning models, using the microphone and the webcam. Users could use their voice and move their face around to affect the sound.
  • An impressive one had six sound models: rain, bouncing ball, sea waves, fire, wind and explosions. It also had a website where each synthesised sound could be compared against real recordings. We couldn’t always tell which was real and which was synthesised!

SoundSelect.png

  • An auditory model of a London Underground train, from the perspective of a passenger on a train, or waiting at a platform. It had a great animation.

train

  • Two projects involved creating interactive soundscapes auralising an image. One involved a famous photo taken by the photographer, Gregory Crewdson. encapsulating  a dark side of suburban America through surreal, cinematic imagery. The other was an estate area, where there are no bodies visible , giving the impression of an eerie atmosphere where background noises and small sounds are given prominence.

And from the Digital Audio Effects class;

  • A create-your-own distortion effect, where the user can interactively modify the wave shaping curve.
  • Input dependant modulation signal based on the physical mass/ spring system
  • A Swedish death metal guitar effect combining lots of effects for a very distinctive sound
  • A very creative all-in-one audio toy, ‘Ring delay’. This  augmented ping-pong delay effect gives controls over the panning of the delays, the equalization of the audio input and delays, and the output gain. Delays can be played backwards, and the output can be set out-of-phase. Finally, a ring modulator can modulate the audio input to create new sounds to be delayed.
  • Chordify, which transforms an incoming signal, ideally individual notes, into a chord of three different pitches.

chordify

  • An audio effects chain inspired by interning at a local radio station. The student helped the owner produce tracks using effects chain presets. But this producers understanding of compressors, EQ, distortion effects… was fairly limited. So the student recreated one of the effects chains into a plugin that only has two adjustable parameters which control multiple parameters inside. 
  • Old Styler, a plug-in that applies sort of a ‘vintage’ effect so that it sounds like from an old radio or an old, black and white movie. Here’s how it sounds.
  • There were some advanced reverbs, including a VST implementation of a state-of-the-art reverberation algorithm known as a Scattering Delay Network (SDN), and a Church reverb incorporating some additional effects to get that ‘church sound’ just right.
  • A pretty amazing cave simulator, with both reverb and random water droplet sounds as part of the VST plug-in.

CaveSimulator

  • A bit crusher, which also had noise, downsampling and filtering to allow lots of ways to degrade the signal.
  • A VST implementation of the Euclidian Algorithm for world rhythms as described by Goddfried Toussaint in his paper The Euclidean Algorithm Generates Traditional Musical Rhythms.
  • A mid/side processor, with excellent analysis to verify that the student got the implementation just right.
  • Multi-functional distortion pedal. Guitarists often compose music in their bedroom and would benefit from having an effect to facilitate filling the song with a range of sounds, traditionally belonging to other instruments. That’s what this plug-in did, using a lot of clever tricks to widen the soundstage of the guitar.
  • Related to the multi-functional distortion, two students created multiband distortion effects.
  • A Python project that separates a track into harmonic, percussive, and residual components which can be adjusted individually.
  • An effect that attempts to resynthesise any audio input with sine wave oscillators that take their frequencies from the well-tempered scale. This goes far beyond auto-tune, yet can be quite subtle.
  • A source separator plug-in based on Dan Barry’s ADRESS algorithm, described here and here. Along with Mikel Gainza, Dan Barry cofounded the company Sonic Ladder, which released the successful software Riffstation, based on their research.

There were many other interesting assignments, including several variations on tape emulation. But this selection really shows both the talent of the students and the possibilities to create new and interesting sounds.

What’s up with the Web Audio API?

Recently, we’ve been doing a lot of audio development for applications running in the browser, like with the procedural audio and sound synthesis system FXive, or the Web Audio Evaluation Tool (WAET). The Web Audio API is part of HTML5 and its a high level Application Programming Interface with a lot of built-in functions for processing and generating sound. The idea is that its what you need to have any audio application (audio effects, virtual instruments, editing and analysis tools…) running as javascript in a web browser.

It uses a dataflow model like LabView and media-focused languages like Max/MSP, Pure Data and Reaktor,. So you create oscillators, connect them to filters, combine them and then connect that to output to play out the sound. But unlike the others, its not graphical, since you write it as JavaScript like most code that runs client-side on a web browser.

Sounds great, right? And it is. But there were a lot of strange choices that went into the API. They don’t make it unusable or anything like that, but it does sometimes leave you kicking in frustration and thinking the coding would be so much easier if only… Here’s a few of them.

  • There’s no built-in noise signal generator. You can create sine waves, sawtooth waves, square waves… but not noise. Generating audio rate random numbers is built in to pretty much every other audio development environment, and in almost every web audio application I’ve seen, the developers have redone it themselves, with ScriptProcessors, AudioWorklets, buffered noise Classes or methods.
  • The low pass, high pass, low shelving and high shelving filters in the Web Audio API are not the standard first order designs, as taught in signal processing and described in [1, 2] and lots of references within. The low pass and high pass are resonant second order filters, and the shelving filters are the less common alternatives to the first order designs. This is ok for a lot of cases where you are developing a new application with a bit of filtering, but its a major pain if you’re writing a web version of something written in MATLAB, Puredata or lots and lots of other environments where the basic low and high filters are standard first order designs.
  • The oscillators come with a detune property that represents detuning of oscillation in hundredths of a semitone, or cents. I suppose its a nice feature if you are using cents on the interface and dealing with musical intervals. But its the same as changing the frequency parameter and doesn’t save a single line of code. There are other useful parameters which they didn’t give the ability to change, like phase, or the duty rate of a square wave. https://github.com/Flarp/better-oscillator is an alternative implementation that addresses this.
  • The square, sawtooth and triangle waves are not what you think they are. Instead of the triangle wave being a periodic ramp up and ramp down, they are the sum of a few terms in the Fourier series that approximate this. This is nice if you want to avoid aliasing, but wrong for every other use. It took me a long time to figure this out when I tried modulating a signal by a square wave to turn it on and off. Again, https://github.com/Flarp/better-oscillator gives an alternative implementation with the actual waveforms.
  • The API comes with a biquad filter that allows you to create almost arbitrary infinite impulse response filters. But you can’t change the coefficients once its created. So its useless for most web audio applications, which involve some control or interaction.

Despite all that, its pretty amazing. And you can get around all these issues since you can always write your own audio worklets for any audio processing and generation. But you shouldn’t have to.

We’ve published a few papers on the Web Audio API and what you can do with it, so please check them out if you are doing some R&D involving it.

 

[1] J. D. Reiss and A. P. McPherson, “Audio Effects: Theory, Implementation and Application“, CRC Press, 2014.

[2] V. Valimaki and J. D. Reiss, ‘All About Audio Equalization: Solutions and Frontiers,’ Applied Sciences, special issue on Audio Signal Processing, 6 (5), May 2016.

[3] P. Bahadoran, A. Benito, T. Vassallo, J. D. Reiss, FXive: A Web Platform for Procedural Sound Synthesis, Audio Engineering Society Convention 144, May 2018

[4] N. Jillings, Y. Wang, R. Stables and J. D. Reiss, ‘Intelligent audio plugin framework for the Web Audio API,’ Web Audio Conference, London, 2017

[5] N. Jillings, Y. Wang, J. D. Reiss and R. Stables, “JSAP: A Plugin Standard for the Web Audio API with Intelligent Functionality,” 141st Audio Engineering Society Convention, Los Angeles, USA, 2016.

[6] N. Jillings, D. Moffat, B. De Man, J. D. Reiss, R. Stables, ‘Web Audio Evaluation Tool: A framework for subjective assessment of audio,’ 2nd Web Audio Conf., Atlanta, 2016

[7] N. Jillings, B. De Man, D. Moffat and J. D. Reiss, ‘Web Audio Evaluation Tool: A Browser-Based Listening Test Environment,’ Sound and Music Computing (SMC), July 26 – Aug. 1, 2015