Charles Brain Puts NVIDIA's Riva to Work Transcribing Radio with a LimeSDR and GNU Radio
Using a GPU to accelerate the task of accurate speech-to-text transcription, this rig tunes in to FM radio stations with ease.
Amateur radio software developer Charles "G4GUO" Brain has been experimenting with blending the worlds of software-defined radio (SDR) and machine learning, feeding NVIDIA's Riva speech recognition system with FM radio signals from GNU Radio.
Using a LimeSDR open-hardware software-defined radio dongle connected to desktop system, Brain has written a GNU Radio flow graph, which tunes the SDR into an FM radio station, demodulates the signal into audio, then feeds the audio into Riva — NVIDIA's GPU-accelerated speech recognition and transcription model.
The result: an automatic transcript of whatever is transmitted. Providing, at least, you have a powerful enough system. "I tried using it on a [GeForce] RTX 2080 Super [graphics] card," Brain explains, but it gave loads of out of memory errors. So the […] example is running on an RTX 3090 with 24G[B] of memory."
"I am running GNU Radio on one computer talking to another computer running the Riva server via Wi-Fi," Brain adds. "I am only doing this because my big machine is nowhere near my antennas. Voice recognition has come a long way since I last played with it 20 or more years ago. Then, I was using IBM's Dragon Naturally Speaking which I used to control an HF radio running my PC-ALE software."
Brain has so far successfully used the system to transcribe BBC Radio 4 programming as wel as narrowband FM (NBFM) signals on the two-meter band. "I have not tried SSB [Single Side Band] on HF [High-Frequency] yet, that will be the next challenge."
Brain's write-up is available on his blog, alongside a demonstration video, while more information on Riva is available on the NVIDIA website.