This thesis describes the application of a perceptually based spectral subtraction algorithm for the enhancement of non-stationary noise corrupted speech. Through examination of speech enhancement techniques, explanations are given for the choice of magnitude spectral subtraction and how the human auditory system can be modelled for frequency domain speech enhancement. It is discovered, that the cochlea provides the mechanical speech enhancement in the auditory system, through the use of masking. Frequency masking is used in spectral subtraction, to improve the algorithm execution time, and to shape the enhancement process making it sound natural to the ear. A new technique for estimation of background noise is presented, which operates during speech sections as well as pauses. This uses two microphones placed on opposite ends of the cellular handset. Using these, the algorithm determines whether the signal is speech, or noise, by examining the current and next frames presented to each microphone. This allows operation in non-stationary conditions, as the estimation is calculated for each frame, and a speech pause is not required for updating. A voting decision process decides the presence of speech or noise which determines which microphone the estimation is calculated from. The importance of an accurate noise estimate is highlighted with a new technique to reduce the effect of musical noise artifacts in the processed speech. This is a classic drawback of spectral subtraction techniques, and it is shown, that the trade off between noise reduction and speech distortion can be extended by this process. A new method for dealing with musical noise is described, which uses a combination of energy and variance examination of the spectrogram to segregate potential musical noise from desired speech sections. By examination of the spectrogram points surrounding musical noise sections, perceptually relevant values replace the corruption leading to cleaner enhanced speech. Any perceptual speech system requires accurate estimates of the clean speech masking thresholds, to prevent noisy sections being passed through the enhancement untouched. In this thesis, a method for the calculation of the estimated clean speech masking thresholds is derived. Classically, this requires an estimation of the clean speech before the thresholds can be derived, but this results in inaccuracy due to the presence of musical noise and spectral nulls. The new algorithm examines the thresholds produced by the corrupted speech, and the background noise, and from these determines the relationship between the two, to produce an estimate of the clean thresholds, with no operation performed on the actual speech signal. A discrepancy is found between the results for male and female speech, which, by examination of the perceptual process, is shown to be due to the different formant positions in male and female speech. Following the development of these parts, the entire enhancement algorithm is tested on a range of noise scenarios, using male and female speech. The results show, that the proposed algorithm is able to provide adequate performance in terms of noise reduction and speech quality.
| Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:561765 |
| Date | January 2000 |
| Creators | Tuffy, Mark |
| Contributors | Laurenson, David |
| Publisher | University of Edinburgh |
| Source Sets | Ethos UK |
| Detected Language | English |
| Type | Electronic Thesis or Dissertation |
| Source | http://hdl.handle.net/1842/1379 |
Page generated in 0.0126 seconds