Model-based noise robustness schemes have been shown to yield excellent performance for automatic speech recognition (ASR) systems, even in low signal-to-noise-ratio (SNR) conditions. These approaches estimate models of the background noise, including both additive and convolutional distortions, and use these to alter the acoustic model parameters to reflect those present in the target environment. These approaches can also be used within an adaptive training environment allowing noise corrupted data to be used efficiently during training to obtain a neutral, canonical, speech model which is suited for adaptation to a range of target environments. This work will initially apply and investigate state-of-the-art model-based compensation approaches developed for ASR systems to keyword spotting in the RATS target domain where SNRs of less than 10dB are specified. This will include improving existing work on discriminative adaptive training approaches based on schemes such as Vector Taylor Series compensation (VTS), Joint Uncertainty Decoding (JUD) and Predictive CMLLR (PCMLLR). In addition, novel forms of model-based compensation specifically aimed at addressing the low SNR environments within the RATS domain will be developed.

One problem with model-based approaches is that they require a representation of how the background noise conditions affect the speech. Approximations in this "mismatch" function can impact performance. Furthermore, though model-based compensation schemes are able to handle background acoustic environments, to achieve the levels of keyword spotting performance required under the RATS programme it will be necessary to adapt, in a fully automated and unsupervised fashion, the acoustic models to be representative of the specific speaker for that utterance. Schemes for combining model-based compensation approaches with speaker adaptation approaches, such as MLLR and CMLLR, will be examined. For these speaker adaptation schemes there is no mismatch function with its associated approximations, general transformations of the acoustic models are estimated. These transformations require more data to obtain robust estimates, and depending on the amount of available data may not handle the non-linearities associated with the impact of background noise conditions. Thus appropriate schemes for combining speaker adaptation approaches with model compensation schemes should yield significant gains. Another challenge is to refine existing speaker adaptation approaches to operate well in low SNR environments. This will build on existing work such as Noisy CMLLR (NCMLLR) adaptation which combines attributes of both model compensation schemes and speaker adaptation. For these low SNR conditions it may be useful to examine discriminative approaches to estimating the transforms. As any hypotheses used to estimate the transform are liable to be error-full, schemes based on discriminative mapping functions will be examined.

Publications (and related papers)