John Ferguson – CEO
Deepwave Digital has open sourced our Radio Automatic Speech Recognition (Radio-ASR) project. By open sourcing this software, we hope that the community will leverage and improve upon it to further its utility for over-the-air speech-to-text recognition. We encourage the open-source community to expand upon this to integrate more complex signal types and incorporate the latest ASR models.
Here is a link to Radio-ASR on our GitHub page.
Back in 2021, I teamed up with Adam Thompson from NVIDIA to present a machine learning + GNU Radio demonstration for the 2021 GNU Radio Conference. You can view that 2021 presentation further down on this page.
For this demonstration, we leveraged the digital signal processing capabilities of GNU Radio to demodulate an over-the-air radio frequency (RF) signal. The current FM demodulator works with the General Mobile Radio Service (GMRS) signal standard to obtain the voice audio stream at a 16 kHz audio rate.
What we did next was to leverage NVIDIA NeMo™ framework to perform automatic speech recognition on the live audio stream. NeMo™ is a toolkit for building and training natural language processing applications that uses a modular approach to speech processing. This allows users to integrate separate functions as needed for their application.
For Radio-ASR, all of this processing occurs in an untethered environment, meaning that the entire processing chain is happening at the edge without the need for an internet connection. This is only possible because Deepwave’s radio hardware incorporates NVIDIA graphics processing units (GPUs) into the signal processing chain. The flow of data between the heterogeneous computing architecture built into our Edge Series AI Radio is shown in the diagram below.
The redeeming feature of Deepwave’s AI Radios is that our architecture eliminates extra memory copies between the GPU and CPU, thereby making real-time software like Radio-ASR possible without the need to send data to an auxiliary laptop or up to the cloud for computing. All processing occurs on the radio.
The Deepwave Radio-ASR system is completely modular. Individual components can be exchanged or customized, for example, by incorporating other demodulation schemes, or language models for languages other than English. This application highlights the key benefits of GPU integration into SDR: the ability to easily leverage pre-trained AI models for a multitude of applications. Read more about GPU integration in to SDRs here.
For more information contact Deepwave Digital: https://deepwavedigital.com/inquiry