This page is an introduction to our Radio Autenocoder (RADE) technology running over baseband FM (BBFM) radios. It has a brief description of the technology and speech samples demonstrating the system. The application is land mobile radio (LMR), for example VHF/UHF walkie talkies, and two-way radios in vehicles for applications such as Amateur Radio, public safety, mining, and recreational vehicles.

Existing protocols for LMR include P.25, Tetra, DMR, M17, DStar and C4FM. Good old analog FM also remains popular as it’s speech quality is competitive with the digital systems, can be used without any infrastructure, and is low cost.

While a variety of LMR protocols have been developed and indeed continue to emerge, the speech quality and physical layer performance of LMR systems hasn’t evolved since the mid 1990s.

Our goal is to present a modern, machine learning (ML) approach for the LMR physical layer, using open source software. This can be considered a drop in replacement for the classical vocoder/FEC signal processing on radios using the BBFM architecture, and can be applied to existing or emerging LMR protocols.

Credits

The RADE BBFM concept evolved from a discussions between Tibor Bece, George Karan, and David Rowe in mid 2024. It is based on the core encoder and decoder from the HF RADE system developed by Jean-Marc Valin and David Rowe. David and Tibor worked through 2024 and 2025 to develop a practical RADE-BBFM system.

The FreeDV Project Leadership Team encouraged this work and David’s work was generously supported by a grant from Amateur Radio Digital Communications (ARDC).

Resources

1. The Radio Autoencoder GitHub repo dr-bbfm branch (released under the two clause BSD license).
2. The development of RADE-BBFM has been documented in a series of blog posts (search on BBFM).
3. Paper: “RADE for Land Mobile Radio: A Neural Codec for Transmission of Speech over Baseband FM Radio Channels”, David Rowe and Tibor Bece, 2025. Available at arxiv.org.

RADE over BBFM Introduction

Consider a modern digital LMR radio (Fig 1). A speech encoder (e.g. AMBE or Codec 2) converts analog speech to a stream of bits, which are the protected from bit errors using forward error correction (FEC) encoder. The bits are shaped with a Tx filter to provide a smoothed stream of symbols that drives a FM modulator. The symbols are sent through what is essentially an analog FM radio (blue squiggle) to the FM demodulator. On the receiver side the opposite process occurs and speech is played out of the radio speaker.

We describe the core analog FM radio signal processing as baseband FM (BBFM). It’s common to analog FM radios and many digital radios. It includes the FM modulator and demodulator, PA, LNA, frequency shifting, RF front end, IF filtering etc. For digital radios we send a sequence of symbols through the BBFM channel, for analog FM we drive it directly with analog speech. The BBFM architecture is popular as it’s low cost, generates a good clean signal, can use low cost non-linear power amplifiers, and low cost crystals for frequency references.

Now consider Fig 2. We have replaced the classical DSP components (speech encoder and decoder, FEC) with Radio Autoencoder (RADE) technology. We keep the baseline BBFM radio, which means this will run on any analog or existing digital FM radio hardware. The advantage is greatly improved speech quality, sensitivity and robustness to fading.

The symbols z that we feed into the FM modulator are similar to the 2 or 4 level symbols from classical digital LMR systems, except they are “analog” in the sense that they have no fixed levels. They could be any value in the range of +/-1, a little like analog FM.

One disadvantage of RADE is the need for more CPU and memory than legacy protocols, which were developed for the constraints of the processors available in the 1990s. However the resources required for RADE are a fraction of that available on a modern laptop or mobile (cell) phone, and likely to fall further with more development.

Demonstration

Here are some samples where we compare analog FM, DStar, and RADE at -100dBm. DStar is representative of a LMR system based on classical DSP vocoder and FEC technology. First, a nice strong -100 dBm signal with no fading which represents the best quality each waveform is capable of delivering.

Now lets try a tough fading channel, that simulates a vehicle moving at 60 km/hr with a weak signal:

The FM and RADE signals levels were -120 dBm, however we had to run DStar at -115 dBm as there was no audio at -120 dBm. DStar did produce output audio down to -121 dBm with no fading. RADE is only mildly affected by the fading, as per our simulation results, and we can even receive intelligible speech down to -127 dBm (!).

ASR Simulation Results

An important goal of our project is improved speech intelligibility over FM. To evaluate intelligibility we need an objective, controlled way of comparing performance. For speech systems this generally means ITU-T P.800 or P.808 standards based subjective testing. However this is complex and requires skills, experience and resources not available to our team.

So for testing RADE we have used automatic speech recognition (ASR) as an objective measure. The general idea is to take a dataset of speech samples, pass them through simulations of analog FM and RADE over BBFM channels, then use a ASR engine to detect the words in the received speech. A post processing system then compares the detected words to the original words and determines the Word Error rate (WER) as a performance metric. Our work uses 500 samples from the Librispeech dataset, and the Whisper ASR system. The Librispeech sentences are complex English sentences, spoken quickly with no contextual cues, and can be difficult to understand even for native English speakers.

A low word error rate (WER) is the goal. The clean (input speech), FARGAN vocoder alone, and Codec 2 3200 (representing a classical DSP vocoder) are provided as controls. The controls are not subject to noise, so are plotted as straight lines.

So green is RADE, blue analog FM. Solid is AWGN (just noise without fading), dotted is a fading channel (a vehicle moving at 60 km/hr). You can compare the difference in dB by drawing a horizontal line at a given WER (e.g. 5%). These curves indicate some impressive results. Compared to analog FM, RADE seems to shrug off multipath fading. It approaches the WER of the baseline FARGAN vocoder at high SNRs.