6G AI/ML Physical Layer: Part 2 - JSCM-Based Audio Traffic

Wireless communication has long been designed with bit accuracy as its top priority, but for services humans directly listen to, like speech, not every bit needs to be perfect. Since preserving meaning is often enough, semantic communication, which shifts the transmission goal from bit-perfect to meaning-preserving, is emerging as a key alternative.

Audio is the most natural target to start this approach with. Speech is the most fundamental means of human communication while also being latency-sensitive, and it exposes the problem that traditional communication systems over-protect bit fidelity beyond what humans actually perceive as quality. This is especially useful in weak-link environments such as satellite access, wearables, and deep fading, or in resource-constrained situations like disaster and public-safety networks.

The core of this approach lies in abandoning the traditional separated source coding - channel coding - modulation structure and combining it into a single end-to-end neural architecture. The transmitter network compresses the input audio and directly produces complex-valued symbols, while the receiver network reconstructs the audio from the noisy symbols. In other words, it is a JSCM (joint source-channel coding and modulation) structure that incorporates the channel itself into the learning process, jointly optimizing compression and transmission.

Evaluation proceeded along two tracks. First, performance was compared against the existing separated pipeline in a 5G-based link-level simulator. Then, a hardware proof-of-concept was built using USRPs in a Bluetooth-band environment to verify operation over an actual wireless link. Both experiments used the LibriSpeech dataset, with comparisons made under fair resource conditions.

The results were clear. SAC-JSCM maintained similar perceptual quality at roughly 10 dB lower SNR or transmit power than the existing benchmark, in both simulation and real hardware. In simulation, it showed a better curve across the entire test range, and in the hardware PoC, while the traditional method exhibited a cliff effect where performance collapsed sharply as power weakened, the semantic method degraded more gradually and held up until the end.

What these results imply is not simply a codec improvement. Future wireless systems could be optimized based on the quality users actually perceive rather than the bitstream itself. Semantic audio communication, starting from speech, is presented as an early example that could lead to lower power consumption, better coverage, and more robust disaster communication.