Pushing boundaries in AI-powered speech enhancement

We recently launched one of our most revolutionary products in history: Infinio Sphere—the first-ever hearing device with a dedicated AI chip. This moment marked a groundbreaking achievement in both the acoustic community and our market. It wasn’t just another product release; we literally made history.

With the overwhelmingly positive feedback we’ve received from launch events, we couldn’t be prouder of our accomplishment. But this success didn’t happen overnight—so, how did we get here?

I believe there are two key factors behind the success of Phonak’s AI-powered hearing device:

1. Deep domain knowledge in tailoring hardware

We have deep domain expertise in designing specific, low-consumption chips that meet the demanding needs of hearing devices. Our hardware is not only powerful but works within the tight constraints imposed by battery usage, memory, and processing cycles.

2. Advanced algorithms backed by massive data training

On the software side, our new noise-cancelling algorithm achieves exceptional performance thanks to extensive training:

• The algorithm was trained on over 22 million audio samples, equivalent to 3 years of audio data!
  
  
• This training was made possible by our competitive AI infrastructure, which includes:
  • 150 GPUs
  • 1300 CPU cores
  • 13,500 GB of RAM

Did you know that the energy required to power this infrastructure—about 1.44 GWh/year—could power 300 European households?

Not just an incremental improvement, a technological leap

A famous quote by author and forward-thinker Oren Harari comes to mind: “The electric light did not come from the continuous improvement of candles.”

Similarly, Infinio Sphere isn’t just an incremental improvement over previous hearing devices – it’s a true technological leap. Our latest product is the result of years of expert knowledge and dedication.

The road to Infinio Sphere

Just in March this year, the paper ‘Sixty Years of Frequency-Domain Monaural Speech Enhancement: From Traditional to Deep Learning Methods’ was published.¹ The authors reviewed the progress made in speech enhancement technology over the last six decades, highlighting the sophistication of modern algorithms.

Despite the advancements, the paper concluded that while deep neural networks (DNNs) are the future of speech enhancement, it’s still technically impossible to run DNNs in real-time on a hearing device due to challenges such as:

• Power consumption
• Memory limitations
• Processing speed

But that’s exactly what we’ve achieved with Infinio Sphere—and we did it well.

AI: The future of hearing care

AI is rapidly becoming an integral part of our lives, and at Sonova, we are ready to embrace it. We are continuously upgrading and extending both our AI infrastructure and domain knowledge to lead the way in hearing care innovation.

Infinio Sphere was the first step forward but is clearly not the destination. This new AI technology is revealing holding an incredible potential for improving our devices from all aspects.

Looking ahead: The sound of the future

I can easily envision a future where everything, from the hardware to the software, is fully optimized by dedicated AI models. The possibilities are limitless, and whatever the future holds, one thing is certain: it will sound amazing.

---

References:

1.Zheng, C., Zhang, H., Liu, W., et al. (2023). Sixty Years of Frequency-Domain Monaural Speech Enhancement: From Traditional to Deep Learning Methods. Trends in Hearing; 27. doi:10.1177/23312165231209913