- Captures S16LE mono audio from the microphone via PulseAudio
- Applies a user-configurable mic gain for input level control
- If the capture device isn't natively 16 kHz, a streaming linear-interpolation resampler converts to 16 kHz (the rate LPCNet expects)
- Processes audio in 10 ms frames (160 samples at 16 kHz)
- For each frame,
lpcnet_compute_single_frame_features()extracts a vector ofNB_TOTAL_FEATURES(36) values:- 18 Bark-scale cepstral coefficients — compact spectral envelope representation
- Pitch period and pitch correlation — fundamental frequency information
- Other auxiliary features used by the downstream neural encoder
- 12 consecutive feature frames (120 ms of speech) are accumulated into a single modem frame buffer of 432 features (12 × 36)
- Once a full modem frame (432 features) is accumulated, it is passed to
rade_tx() - The RADE encoder is a neural network (autoencoder) that compresses the 432-dimensional feature vector into a sequence of OFDM-like complex symbols (960 IQ samples at 8 kHz)
- The encoding is designed to be robust to HF radio channel impairments — multipath fading, frequency offset, noise — by learning a channel-resilient latent representation during training
- The complex output represents a baseband waveform ready for transmission
- An optional bandpass filter (centre 1600 Hz, bandwidth 1500 Hz) constrains the output spectrum to the SSB passband (~700–2300 Hz)
- Prevents out-of-band energy from causing interference on adjacent frequencies
- Can be toggled at runtime via
set_bpf_enabled()
- The real part of the complex IQ signal is extracted and scaled by a configurable TX scale factor (default 16384)
- Resampled from the 8 kHz modem rate to the output device's native rate if needed
- Written to the radio soundcard via PulseAudio for transmission as an SSB audio signal
- When transmission stops, a special end-of-over frame (1152 samples) is sent via
rade_tx_eoo() - This signals to the receiver that the transmission is complete, allowing clean decoder shutdown
- Input/output level metering: RMS levels are computed and exposed via thread-safe atomics for UI display
- TX spectrum display: A 512-point FFT with a Hann window computes the power spectrum of the TX output, available via
get_spectrum()for real-time visualisation - Prefill buffer: 2 modem frames of silence are written to the output buffer at startup to absorb the ~120 ms latency of accumulating 12 feature frames before the first output is produced
| Parameter | Value |
|---|---|
| Feature frame | 10 ms (160 samples @ 16 kHz) |
| Modem frame | 120 ms (12 feature frames) |
| TX output per modem frame | 960 IQ samples @ 8 kHz |
| End-of-over frame | 1152 IQ samples @ 8 kHz |
| Algorithmic latency | ~120 ms (accumulation) + audio buffering |
- RADE is part of the Codec 2 project by David Rowe (VK5DGR)
- Built on top of the Opus/LPCNet neural speech processing framework
- Captures S16LE mono audio from the microphone via PulseAudio
- Applies a user-configurable mic gain for input level control
- If the capture device isn't natively 16 kHz, a streaming linear-interpolation resampler converts to 16 kHz (the rate LPCNet expects)
- Processes audio in 10 ms frames (160 samples at 16 kHz)
- For each frame,
lpcnet_compute_single_frame_features()extracts a vector ofNB_TOTAL_FEATURES(36) values:- 18 Bark-scale cepstral coefficients — compact spectral envelope representation
- Pitch period and pitch correlation — fundamental frequency information
- Other auxiliary features used by the downstream neural encoder
- 12 consecutive feature frames (120 ms of speech) are accumulated into a single modem frame buffer of 432 features (12 × 36)
- Once a full modem frame (432 features) is accumulated, it is passed to
rade_tx() - The RADE encoder is a neural network (autoencoder) that compresses the 432-dimensional feature vector into a sequence of OFDM-like complex symbols (960 IQ samples at 8 kHz)
- The encoding is designed to be robust to HF radio channel impairments — multipath fading, frequency offset, noise — by learning a channel-resilient latent representation during training
- The complex output represents a baseband waveform ready for transmission
- An optional bandpass filter (centre 1600 Hz, bandwidth 1500 Hz) constrains the output spectrum to the SSB passband (~700–2300 Hz)
- Prevents out-of-band energy from causing interference on adjacent frequencies
- Can be toggled at runtime via
set_bpf_enabled()
- The real part of the complex IQ signal is extracted and scaled by a configurable TX scale factor (default 16384)
- Resampled from the 8 kHz modem rate to the output device's native rate if needed
- Written to the radio soundcard via PulseAudio for transmission as an SSB audio signal
- When transmission stops, a special end-of-over frame (1152 samples) is sent via
rade_tx_eoo() - This signals to the receiver that the transmission is complete, allowing clean decoder shutdown
- Input/output level metering: RMS levels are computed and exposed via thread-safe atomics for UI display
- TX spectrum display: A 512-point FFT with a Hann window computes the power spectrum of the TX output, available via
get_spectrum()for real-time visualisation - Prefill buffer: 2 modem frames of silence are written to the output buffer at startup to absorb the ~120 ms latency of accumulating 12 feature frames before the first output is produced
| Parameter | Value |
|---|---|
| Feature frame | 10 ms (160 samples @ 16 kHz) |
| Modem frame | 120 ms (12 feature frames) |
| TX output per modem frame | 960 IQ samples @ 8 kHz |
| End-of-over frame | 1152 IQ samples @ 8 kHz |
| Algorithmic latency | ~120 ms (accumulation) + audio buffering |
- RADE is part of the Codec 2 project by David Rowe (VK5DGR)
- Built on top of the Opus/LPCNet neural speech processing framework
LPCNet is a neural network-based speech synthesis model developed by Jean-Marc Valin (Mozilla/Xiph.org), designed for low-complexity, real-time speech coding.
It combines two classic techniques:
-
Linear Predictive Coding (LPC) — a traditional DSP method that models the vocal tract as an all-pole filter. LPC captures the spectral envelope (formants) of speech very efficiently.
-
WaveRNN — a recurrent neural network that generates audio samples one at a time.
The key insight is: instead of having the neural network learn everything about the speech signal, let the LPC filter handle the predictable spectral structure, and have the neural network only model the residual excitation signal. This dramatically reduces the complexity of what the network needs to learn.
- Frame-rate network: Processes features (Bark-scale cepstral coefficients, pitch period, pitch correlation) once per frame (~10ms), producing a conditioning vector.
- Sample-rate network: A lightweight GRU that generates one audio sample at a time, conditioned on the frame-rate output and the LPC prediction. It outputs a probability distribution over sample values, using a dual-softmax formulation.
- Runs in real-time on a single CPU core (no GPU needed)
- Achieves near-transparent speech quality at very low bitrates (~1.6 kbps when paired with a codec like Codec 2)
- The LPC "shortcut" reduces the neural network's workload by ~30dB, making the tiny network feasible
This project uses LPCNet (or a derivative like RADE — Radio Autoencoder) for encoding/decoding speech over radio channels, where robustness to channel errors and low bitrate are critical.
- Original paper: "LPCNet: Improving Neural Speech Synthesis Through Linear Prediction" (Valin & Skoglund, 2019, ICASSP)
FARGAN is a neural vocoder developed by Jean-Marc Valin as a successor to LPCNet, designed for ultra-low complexity speech synthesis while maintaining high audio quality.
LPCNet generates audio one sample at a time, which creates a sequential bottleneck. Even though it's efficient for a neural vocoder, sample-level autoregression limits parallelism and throughput. FARGAN addresses this by operating at the frame level instead.
Rather than predicting individual samples, FARGAN generates an entire frame of audio (~10ms / 160 samples at 16kHz) in one step, using a GAN (Generative Adversarial Network) training framework. It retains the LPC filtering concept from LPCNet to reduce what the network must learn.
- Frame-level autoregression: The model is autoregressive across frames, not samples. Each forward pass produces a full frame of audio, conditioned on previous frames.
- LPC synthesis filter: Like LPCNet, an LPC filter models the vocal tract. The neural network generates the excitation signal, which is then passed through the LPC filter to produce the final waveform.
- Sub-frame structure: Each frame is broken into sub-frames, with a lightweight recurrent structure operating across sub-frames for temporal coherence.
- GAN training: A discriminator network is used during training to push the generator toward producing natural-sounding speech (adversarial loss), combined with feature matching and spectral losses.
| Feature | LPCNet | FARGAN |
|---|---|---|
| Generation unit | 1 sample | 1 frame (~160 samples) |
| Autoregression | Sample-level | Frame-level |
| Training | Cross-entropy | GAN (adversarial) |
| Complexity | ~3 GFLOPS | ~0.5 GFLOPS |
| Parallelism | Minimal (sequential samples) | High (within-frame parallel) |
- ~6x less compute than LPCNet while achieving comparable or better quality
- Easily runs in real-time on low-power devices (embedded, mobile)
- Well-suited for pairing with low-bitrate speech codecs (e.g., Codec 2, RADE) where the vocoder reconstructs speech from compact feature vectors
- GAN training produces crisper, more natural output than the probability-based sample generation in LPCNet
In the RADE (Radio Autoencoder) pipeline, FARGAN serves as the decoder-side vocoder — it takes the decoded feature vectors (Bark-scale cepstral coefficients, pitch parameters) received over the radio channel and synthesizes them back into audible speech. Its low complexity makes it practical for real-time amateur radio applications.
- "FARGAN: Efficient Neural Audio Synthesis with Framewise Auto-Regressive GAN" (Valin, 2024)
- Built on the Opus/Codec 2 ecosystem from Xiph.org