Abstract
Automatic Speech Recognition (ASR) in vehicular environments is hindered by diverse noise sources, including engine hum, wind, and transient interference. This paper proposes a robust front-end system using a two-microphone array that integrates adaptive null-forming with auditory subbanding based on the Bark scale. By decomposing signals into non-uniform frequency bands, the system can steer "nulls" independently across the spectrum, effectively isolating speech from complex noise. Experimental results in real driving scenarios demonstrate significant improvements in Signal-to-Noise Ratio (SNR) and ASR accuracy compared to traditional methods.
Introduction
The automotive cabin presents a hostile environment for speech recognition systems. Engineers must contend with persistent low-frequency engine noise, high-frequency wind whistle, and unpredictable transient sounds like passing traffic or radio broadcasts. While large microphone arrays can use beamforming to focus on a driver's voice, the physical and cost constraints of vehicle dashboards often limit systems to just two microphones. With such a limited array, creating a narrow enough beam to isolate speech is challenging. This research shifts the focus from beamforming to adaptive "null-forming" — digitally creating a "cone of silence" and aiming it at noise sources to enhance the target speech signal.
Methodology
The proposed system architecture consists of three primary stages:
- Adaptive Null-Forming: Utilizing a two-microphone "end-fire" configuration (where one microphone is placed behind the other), the system calculates signal differences to cancel out sounds arriving from specific angles.
- Auditory Subbanding: Unlike standard wideband null-forming, this system employs a non-uniform filterbank based on the Bark scale, which mimics human auditory perception. Low-frequency bands are narrower to capture dense speech energy, while high-frequency bands are broader. This allows the system to steer the "null" to different directions for different frequencies, a critical capability when multiple noise sources are present.
- Single-Channel Enhancement: A final cleanup phase uses Wiener filtering and minimum statistics noise estimation to remove residual hiss and stationary noise.
Experimental Setup
The system was evaluated through both simulations and real-world vehicle tests. The real-vehicle testing involved recording 350 Chinese phrases under three distinct conditions:
- Case 1: Driving at 75 km/h with windows closed (dominated by engine and road noise).
- Case 2: Parked roadside with windows open (ambient environmental noise).
- Case 3: Parked with the radio playing at high volume (extreme interference).
Key Results
The auditory subband approach consistently outperformed standard single-band methods, providing a 2–3 dB improvement in SNR. The impact on ASR accuracy was particularly notable across the tested scenarios.
In Case 1, the recognition rate improved from 64.3% to 77.3%. In the quietest scenario (Case 2), accuracy reached 88.6%. Even in the extreme interference of Case 3, the system nearly doubled the recognition rate from 15.2% to 29.0%, although this remains a challenging area for current technology.
Discussion and Conclusion
The primary innovation of this work is the use of frequency-dependent nulls. By splitting the signal into auditory subbands, the system can effectively "multi-task," canceling different noise types across the frequency spectrum simultaneously. While the system showed limitations in handling loud, speech-like interference (such as radio broadcasts in Case 3), it proves that a high-performance ASR front-end can be achieved with a minimal two-microphone array through bio-mimetic signal processing. Future work will focus on improving the integration of null-forming with post-filtering to better handle transient, non-stationary interferences.
Summary updated