Happening on-device with increasingly integrated silicon, the use of artificial intelligence in digital audio processing is having massive implications to the audio industry. Read all about it in the November 2021 edition of audioXpress (I know, we released the October edition just two weeks ago).

This year, given the circumstances of the extended global pandemic effects, writing a DSP report was even harder. Usually, we would be able to work through a series of documents submitted by manufacturers and key vendors, some collected meticulously in the field during industry events over the past 12 months, others anticipating whatever these companies intended to release in the next trade-show or industry conference.

Such events are now a distant memory. While I see that InfoComm has not been cancelled yet - which is almost shocking under the serious circumstances of the pandemic in the US (and particularly Florida) - any chance to visit the always-secretive demo rooms of these companies is not going to happen before January, or even mid-2022... And while online interactions or attending webinars is a possibility, it's hardly a rewarding experience when you need to collect factual information that you can report to readers - unless you are focused on just one thing all the time - as some of my fortunate colleagues are.

Nevertheless, I believe the report that you will be able to read in audioXpress November 2021 - much like the ones published for the past two years - will be worth the subscription cost of this magazine. This edition was also reinforced with two great contributed pieces, one covering DSP applications in automotive interior acoustics (an article by Roger Shively) and another on the inspiring potential of an integrated DSP solution for home listening and high-end audio applications (written by Al Clark). I take this opportunity to formally invite more industry experts reading this newsletter to submit contributed pieces on the diverse topics we need to address every month, focused on audio development. (Email)

But I didn't mean to write this newsletter just to promote the November edition of audioXpress. I am writing to highlight the fact that the use of AI and ML to determine digital signal processing of audio is something that all manufacturers, in all application fields, need to take into account. Because that is my main takeaway from the research I did recently. There's real progress and things are moving fast. Much faster than you think (and business is there - there's money to be made).

Unlike progress in many other important audio research areas, which product developers and product managers will all tell you to forget about because of the law of diminishing returns (unless you work in high-end audio and price is no obstacle), the use of AI and ML in audio can be more disruptive than in any other field, meeting real market requirements and consumers' needs. And that's the reason why we are seeing so many advancements and announcements.

Syntiant just announced its latest NDP120 Neural Decision Processor (NDP) chip for audio and sensor processing in battery-powered devices. The low-power Syntiant Core 2 technology contained in this latest chip delivers 25x the power of NDP devices currently shipping, supporting echo-cancellation, beamforming, noise suppression, speech enhancement, speaker identification, keyword spotting, multiple wake words, event detection, and local commands recognition. At the same time Cadence now offers decoder implementations supporting MPEG-H 3D Audio to designs using its Tensilica HiFi DSPs, opening up the potential to bring interactive and immersive sound to millions of users worldwide.

Just among the news published this week at audioXpress.com, we find important examples, with crucial announcements of strategic hardware and software partnerships, new development platforms, and new silicon being released. The full audio DSP connection between those announcements is often missing but the actual applications are clear as AI/ML moves to "on-device" - the current industry buzzword is “edge AI.” Which in audio terms means making things happen closer to the microphone, or even directly in the ear. Obvious references always include voice recognition applications, and voice processing for speech optimizations, as well as active noise cancellation (ANC), but current hot trends are adding spatial audio processing and audio enhancement for augmentation, as well.

In the report I wrote for audioXpress, because of page limitations, I was forced to leave out numerous practical application examples of companies working in this area, particularly in professional audio. But I was able to browse through some of the most promising examples of audio processing that will be exponentially improved when combined with ML models, and controlled with AI-enabled processes. To have an idea of what's coming we just need to consider the power of such platforms, combined with sensor fusion, geolocation and eventually connected services. All things available now.

Imagine being able to remove reverberation from a live audio input signal - not just fixed frequency-dependent noise. Or having an adaptive ANC engine able to adjust and filter noises or pass-through sounds depending on current activity profiles and user location. Applying effective hearing enhancement and communications using optimized algorithms that have been extensively trained, but are also able to adjust in real time throughout changing acoustic environments, purely based on AI, without the need for user input.

I still remember when I saw the Waves plug-ins that we all used to carefully manage on a multitrack DAW session running on the most powerful workstation we could buy, suddenly being able to run directly on each-channel of a digital console. In professional audio, we are seeing progress advancing with each product iteration and DSP becoming ubiquitous in all classes of products. The most sophisticated studio processing tools are now available for all creators working at home, through cloud-based services and resources, which are understandably pioneering AI/ML for differentiation. Now, consumer audio is moving even faster.

And many of the offline software tools still available for the studio are basically benefiting from highly trained ML models, which very quickly we will be able to port to those edge AI platforms. A few examples can be found in tools such as the latest DeVerberate 3 plug-in from Acon Digital, which is able to attenuate unwanted reverb present in recorded audio. The just-released version includes an entirely new algorithm based on deep learning for fully automatic reverb reduction of recorded voice, even with multiple voices speaking at the same time. By training a neural network on thousands of high-quality voice recordings and a wide variety of acoustical surroundings, the plug-in can automatically distinguish speech from reverb. Sounds familiar? Yes, the same type of processing is actually being applied in smartphones, communication headsets, and even true wireless earbuds with ANC.

Actually, an improved and much more sophisticated version was recently demonstrated by Descript - a company I have mentioned previously in this space, which converts speech to text and allows edit decisions on recorded audio and video content, simply by editing text. Descript is working on sophisticated AI applied to audio processing and they have announced Studio Sound, a powerful new feature that enhances speakers’ voices while reducing and removing background noise, room echo, and other undesired sounds. Studio Sound is still in beta but already shows great potential to separate voice from the recorded acoustic effects, helping anyone recorded over a Zoom call to sound like they are using a broadcast microphone in a properly treated studio. And Descript also offers AI-based ducking, which automatically lowers the volume on music and ambient sounds when speech is present. Amateur podcasters love this.

I also mentioned AI being used for spatial processing. And that in itself is a topic for many other articles. I'll just mention an example from LANDR, the online audio mastering service that uses the power of artificial intelligence and data from more than 10 million professionally mastered music tracks to become a major platform in the music industry. Musicians have been using LANDR to master and submit their work directly to digital streaming platforms, normally in stereo. But now that Apple Music, Tidal, and Amazon Music are promoting Dolby Atmos for music, LANDR is also offering spatial audio mastering services to that file format. Given the “creative” possibilities of mixing music in “spatial audio,” I am not surprised that an AI-based service could do better than stoned “music producers”.

But as I feared, that includes "remastering" stereo audio into Atmos using LANDR AI-based, automated processes. Yes, artificial intelligence will not help much when the idea is stupid in the first place.

audioXpress | Voice Coil | LIS 

audioXpress features great articles, projects, tips, and techniques for the best in quality audio. It connects manufacturers and distributors with audio engineers and enthusiasts eager for innovative solutions in sound, acoustic, and electronics.

Voice Coil, the periodical for the loudspeaker industry, delivers product reviews, company profiles, industry news, and design tips straight to professional audio engineers and manufacturers who have the authority to make powerful purchasing decisions.

The Loudspeaker Industry Sourcebook is the most comprehensive collection of listings on loudspeaker material in the industry. Purchasers and decision makers refer to the guide for an entire year when making selections on drivers, finished systems, adhesives, domes, crossovers, voice coils, and everything in between.