Speech Enhancement
Comparison of Our Algorithm with iPhone4
In the past year, dual mics have been used in iPhone4 and Google's Nexus One. The performance is reasonably good if the background noise is stationary. However, if noise is intermittent or non-stationary, then the noise suppression performance is not good. In addition, the adaptation is also slow even in stationary noisy environment. We performed some experiments to compare our algorithm with that of iPhone4. It was found that the intelligibility is not good in iPhone4 as compared to ours. Here we include 2 sound files, which were recorded in the same environment. One can clearly judge which one is better.
Audio Demos: iPhone Result, Our Result
Dual-Mic Approach to Speech Enhancement in Extremely Noisy Battlefield Environment
When soldiers call for close air support (CAS), they are normally very near the frontline where mortar explosions, machine guns, and other types of background noises are overwhelming. Communication quality may be compromised. Conventional speech enhancement techniques can only deal with stationary noise, but not intermittent and unpredictable noise such as explosions. Moreover, conventional techniques cannot deal with large amplitude noise where the signal-to-noise ratio (SNR) may be less than zero; that is, the voice amplitude is much less than the background noise. We developed a dual-mic approach for speech enhancement. Unlike conventional adaptive filters (LMS, NLMS, RLS) that usually work in time-domain, our novel adaptive algorithm works in the frequency domain and achieves superb sound quality. The following recorded files illustrate the performance of our approach. The SNR in Mic-1 is -20.1837 dB and the SNR in Mic-2 is -22.3265 dB. Both are very low. The speech is not intelligible in such battlefield environment. However, after speech enhancement, we can recover the speech quite clearly (the SNR becomes 22.2848dB). About Recording Setup
Audio Demos: Mic1, Mic2, Result
Minimizing Interactions between Inbound and Outbound Signals in Helmet
NASA is developing a new generation of audio system for astronauts. The inbound signals are played using speakers inside the helmet and the outbound signals are collected using a built-in mic. However, inbound signals may interfere with outbound signals and create an annoying positive feedback during communications. The outbound signal quality is also susceptible to reverberation and structure borne noises. We developed novel adaptive filters to minimize interactions between inbound and outbound signals. In Phase 1, we performed extensive simulations and experiments to prove the feasibility of our approach. Actual data were recorded in helmet. The speech quality was significantly improved from a PESQ score of less than 2 before filtering to more than 3 after filtering. The computational requirement of our method is also minimal, as compared to others. The figures below demonstrate the PESQ score and speed comparison of our method with several other methods. In addition, we include some sound files before and after filtering. The male voice is the outbound signal and the female voice is the inbound signal, which interferes with the outbound signal. After processing, the female voice is completely suppressed using our algorithm whereas other methods still contain residual female voice.
Audio Demo: Original, NLMS, AP, RLS, Our
Applications: Background interference suppression, Echo Cancellation, VoIP