Introduction
In-car communication (ICC) systems are designed to improve speech intelligibility within the vehicle cabin by amplifying the driver's voice for passengers in the rear. However, these systems inherently create a closed-loop acoustic feedback path. Without effective Acoustic Echo Cancellation (AEC), the amplified signal is picked up by the microphone, leading to distracting echoes or even instability.
A critical constraint for ICC systems is latency; to ensure a natural "mouth-to-ear" experience, the total algorithmic delay must typically remain below 12ms. While Frequency Domain Adaptive Filtering (FDAF) is the standard for AEC due to its computational efficiency, its block-based nature often introduces delays that exceed this threshold. Reducing the block size to lower delay usually results in a prohibitive increase in computational complexity. This paper proposes a "delay-flexible" Stereo AEC (SAEC) algorithm that decouples delay from filter adaptation to resolve this trade-off.
Proposed Methodology
The core innovation lies in a modified state-space frequency-domain SAEC architecture. Unlike traditional FDAF, which links the frame shift, window length, and update rate, the proposed system introduces three independent parameters:
- Frame Shift ($R$): Controls the actual algorithmic delay.
- Microphone Input Frame Length ($L$): Acts as a "look-back" buffer, allowing the system to analyze a longer window of audio than the current frame shift.
- Update Rate ($U$): Determines the frequency of filter coefficient recalculation.
By decoupling these parameters, the system can output audio with a very low frame shift (small $R$) while performing the computationally intensive filter updates less frequently (larger $U$). To maintain performance despite infrequent updates, the "look-back" buffer ($L > R$) ensures the adaptive algorithm retains sufficient context to model the car's acoustic environment accurately.
Experimental Setup
The system was evaluated using simulated impulse responses typical of a car cabin (reverberation time $T_{60} = 50$ms). Test signals followed ITU-T Recommendation P.501, simulating challenging scenarios including single-talk, interruptions, and double-talk. The audio source consisted of 10 seconds of FM radio speech followed by 35 seconds of active double-talk.
Key Results and Discussion
The researchers compared the delay-flexible approach against a high-delay baseline ($L=R=U=256$). The results demonstrate that the proposed method can significantly reduce latency with minimal impact on echo suppression performance.
As shown in the table above, the baseline system with a 32ms delay achieved an average Echo Return Loss Enhancement (ERLE) of 14.95 dB. By adjusting the parameters to $L=256, R=32, U=32$, the algorithmic delay was reduced to just 4ms. This configuration achieved an ERLE of 15.85 dB—actually outperforming the baseline in suppression while providing a massive improvement in latency. Furthermore, the system distance ($d_{sys}$), which measures how accurately the algorithm models room acoustics, remained stable across configurations.
Computationally, the proposed method is approximately 45% less expensive than a naive low-delay implementation ($L=R=U=32$), proving that decoupling the update rate is a viable strategy for resource-constrained automotive Digital Signal Processors (DSPs).
Conclusion
This paper presents a flexible SAEC framework that successfully breaks the traditional link between algorithmic delay and computational complexity in frequency-domain AEC. By utilizing a look-back buffer and independent update rates, the system achieves the sub-12ms latency required for high-quality ICC systems without sacrificing echo suppression performance. Future work will focus on extending this model to bidirectional ICC systems and testing on real-world automotive hardware.
Summary updated