The Research And Implementation Of Speech Enhancement Algorithm Based On Bark Domain

In speech communication speakers often are interfered by background noise, and the intense background noise will distort the quality of speech, even words are hardly intelligible. So enhancing the noisy signal by extracting as much as possible the original pure voice, to improve the performance of speech communication, is an important study in speech signal processing. It involves kinds of applications, such as hearing aids, the blind voice communication, human-computer interaction system and mobile voice communication and so on. At present, the voice is the main objects for transmitting in the mobile communications and Voice over Internet Protocol (VOIP) technology, and speech enhancement will occupy more important role in the communication field, and its applications will be more widely in the embedded multimedia platform.There are lots of speech enhancement technologies, and the spectral subtraction based on short-term spectral estimation needing a small amount of computing, enhancing effective, easy to achieve real-time, have been widely concerns and applications in embedded speech signal field. However, under low speech-noise-ratio (SNR) or non-stationary noise conditions, the enhanced voice often accompanied by'the music noise'and the background noise that can not be suppressed, and at the same time a great voice distortion.Based on the traditional spectral subtraction, the paper presents a quickly self-adapting noise spectrum estimation method in bark domain. This new method bases on the perceptual model, transforming the noisy signal from frequency domain to bark domain, estimating noise according to the updated local minimum and signal-presence probability, and then execute speech enhancement based on masking properties in bark domain to improve the voice perceptual effects. Finally, the paper has transplanted the proposed speech enhancement algorithm, into the ADI's new ADSP-BF533 embedded multimedia processor platform, and has achieved the implementation of real-time high-quality audio data acquisition and output, and suppresses noise well.