Mask-based enhancement of very noisy speech
File(s)
Author(s)
Lightburn, Leo Charles
Type
Thesis
Abstract
When speech is contaminated by high levels of additive noise, both its perceptual quality and its intelligibility are reduced. Studies show that conventional approaches to speech enhancement are able to improve quality but not intelligibility. However, in recent years, algorithms that estimate a time-frequency mask from noisy speech using a supervised machine learning approach and then apply this mask to the noisy speech have been shown to be capable of improving intelligibility.
The most direct way of measuring intelligibility is to carry out listening tests with human test subjects. However, in situations where listening tests are impractical and where some additional uncertainty in the results is permissible, for example during the development phase of a speech enhancer, intrusive intelligibility metrics can provide an alternative to listening tests. This thesis begins by outlining a new intrusive intelligibility metric, WSTOI, that is a development of the existing STOI metric. WSTOI improves STOI by weighting the intelligibility contributions of different time-frequency regions with an estimate of their intelligibility content. The prediction accuracies of WSTOI and STOI are compared for a range of noises and noise suppression algorithms and it is found that WSTOI outperforms STOI in all tested conditions.
The thesis then investigates the best choice of mask-estimation algorithm, target mask, and method of applying the estimated mask. A new target mask, the HSWOBM, is proposed that optimises a modified version of WSTOI with a higher frequency resolution. The HSWOBM is optimised for a stochastic noise signal to encourage a mask estimator trained on the HSWOBM to generalise better to unseen noise conditions. A high frequency resolution version of WSTOI is optimised as this gives improvements in predicted quality compared with optimising WSTOI. Of the tested approaches to target mask estimation, the best-performing approach uses a feed-forward neural network with a loss function based on WSTOI. The best-performing feature set is based on the gains produced by a classical speech enhancer and an estimate of the local voiced-speech-plus-noise to noise ratio in different time-frequency regions, which is obtained with the aid of a pitch estimator.
When the estimated target mask is applied in the conventional way, by multiplying the speech by the mask in the time-frequency domain, it can result in speech with very poor perceptual quality. The final chapter of this thesis therefore investigates alternative approaches to applying the estimated mask to the noisy speech, in order to improve both intelligibility and quality. An approach is developed that uses the mask to supply prior information about the speech presence probability to a classical speech enhancer that minimises the expected squared error in the log spectral amplitudes. The proposed end-to-end enhancer outperforms existing algorithms in terms of predicted quality and intelligibility for most noise types.
The most direct way of measuring intelligibility is to carry out listening tests with human test subjects. However, in situations where listening tests are impractical and where some additional uncertainty in the results is permissible, for example during the development phase of a speech enhancer, intrusive intelligibility metrics can provide an alternative to listening tests. This thesis begins by outlining a new intrusive intelligibility metric, WSTOI, that is a development of the existing STOI metric. WSTOI improves STOI by weighting the intelligibility contributions of different time-frequency regions with an estimate of their intelligibility content. The prediction accuracies of WSTOI and STOI are compared for a range of noises and noise suppression algorithms and it is found that WSTOI outperforms STOI in all tested conditions.
The thesis then investigates the best choice of mask-estimation algorithm, target mask, and method of applying the estimated mask. A new target mask, the HSWOBM, is proposed that optimises a modified version of WSTOI with a higher frequency resolution. The HSWOBM is optimised for a stochastic noise signal to encourage a mask estimator trained on the HSWOBM to generalise better to unseen noise conditions. A high frequency resolution version of WSTOI is optimised as this gives improvements in predicted quality compared with optimising WSTOI. Of the tested approaches to target mask estimation, the best-performing approach uses a feed-forward neural network with a loss function based on WSTOI. The best-performing feature set is based on the gains produced by a classical speech enhancer and an estimate of the local voiced-speech-plus-noise to noise ratio in different time-frequency regions, which is obtained with the aid of a pitch estimator.
When the estimated target mask is applied in the conventional way, by multiplying the speech by the mask in the time-frequency domain, it can result in speech with very poor perceptual quality. The final chapter of this thesis therefore investigates alternative approaches to applying the estimated mask to the noisy speech, in order to improve both intelligibility and quality. An approach is developed that uses the mask to supply prior information about the speech presence probability to a classical speech enhancer that minimises the expected squared error in the log spectral amplitudes. The proposed end-to-end enhancer outperforms existing algorithms in terms of predicted quality and intelligibility for most noise types.
Version
Open Access
Date Issued
2019-11
Date Awarded
2020-03
Copyright Statement
Creative Commons Attribution NonCommercial Licence
Advisor
Brookes, David
Sponsor
Engineering and Physical Sciences Research Council
Publisher Department
Electrical and Electronic Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)