Combining Speech Fragment Decoding and Adaptive Noise Floor Modeling

Abstract

Abstract-This paper presents a novel noise-robust automatic speech recognition (ASR) system that combines aspects of the noise modeling and source separation approaches to the problem. The combined approach has been motivated by the observation that the noise backgrounds encountered in everyday listening situations can be roughly characterized as a slowly varying noise floor in which there are embedded a mixture of energetic but unpredictable acoustic events. Our solution combines two complementary techniques. First, an adaptive noise floor model estimates the degree to which high-energy acoustic events are masked by the noise floor (represented by a soft missing data mask). Second, a fragment decoding system attempts to interpret the high-energy regions that are not accounted for by the noise floor model. This component uses models of the target speech to decide whether fragments should be included in the target speech stream or not. Our experiments on the CHiME corpus task show that the combined approach performs significantly better than systems using either the noise model or fragment decoding approach alone, and substantially outperforms multicondition training.Index Terms-Adaptive noise floor modeling, fragment decoding, missing data decoding, noise robust speech recognition.

Sound event recognition in unstructured environments using spectrogram image processing

“…The system then proceeds by searching for the best combination of fragments that represent the target source. This can be further enhanced by utilising knowledge about the noise present in the signal [144]. An alternative approach in [108] uses an uncertainty decoder to allow for errors made in the mask estimation process.…”

A sub-band-based feature reconstruction approach for robust speaker recognition

“…A vector is said to belong to the cluster that is most likely to have generated it. As the distribution of the vector is assumed to be Gaussian, the cluster membership m− → x (t) of a vector − → x (t) is defined as (15) and then the unreliable components of the vector are reconstructed using MAP estimation method.…”

“…In MDT, two different methods have been considered to perform speech or speaker recognition with incomplete data: marginalization [13][14][15] and reconstruction [16,17]. In marginalization, the unreliable components are discarded or integrated up to the observed values.…”

Τεχνικές Αραιής Και Χαμηλής Τάξης Αναπαράστασης Για Εύρωστη Αναγνώριση Ομιλητή Και Ανακατασκευή Ελλιπών Χαρακτηριστικών

“…Imputation [48,49,50,51,52,53,54,55] is dened as the technique of substituting missing time-frequency components with an estimate of the time-frequency component value based on speech signal's high degree of redundancy. In marginalization [56,57,58,59], missing spectrotemporal regions are ignored and thus, recognition is based on the reliable components of the noisy speech signal's time-frequency representation, where observation likelihoods are computed by integrating over the range of possible values of the missing components. All these methods exploit various speech signals properties to estimate the missing features, from the data correlation expressed through statistical models to sparsity-based estimation where the features are sparsely represented in a given dictionary.…”

“…MDT have been extensively applied in the context of robust automatic speech recognition (ASR) as a solution to performance degradation due to noisy speech features, and they are distinguished in two main categories, namely, marginalization and imputation. In marginalization [56,58,59], speech decoding is based on the reliable components of a noisy time-frequency representation, while the unreliable components are eliminated or marginalized up to the observed values. The imputation approach [49,50,51,52,53,54,55] is associated with the estimation of the missing data, so that decoding can be performed in a conventional manner.…”