Mask estimation and imputation methods for missing data speech recognition in a multisource reverberant environment

Keronen, Sami; Kallasjoki, Heikki; Remes, Ulpu; Brown, Guy J.; Gemmeke, Jort F.; Palomäki, Kalle

doi:10.1016/j.csl.2012.06.005

Cited by 12 publications

(1 citation statement)

References 28 publications

(44 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In [128], the authors presented an automatic speech recognition system that uses a missing data approach to compensate for challenging environmental noise containing both additive and convolutive components. The unreliable and noise corrupted "missing" components are identified using a Gaussian mixture model (GMM) classifier based on a diverse range of acoustic features.…”

Section: Missing Feature Techniquesmentioning

confidence: 99%

Acoustic event detection and localization using distributed microphone arrays

Chakraborty

View full text Add to dashboard Cite

Automatic acoustic scene analysis is a complex task that involves several functionalities: detection (time), localization (space), separation, recognition, etc. This thesis focuses on both acoustic event detection (AED) and acoustic source localization (ASL), when several sources may be simultaneously present in a room. In particular, the experimentation work is carried out with a meeting-room scenario. Unlike previous works that either employed models of all possible sound combinations or additionally used video signals, in this thesis, the time overlapping sound problem is tackled by exploiting the signal diversity that results from the usage of multiple microphone array beamformers. The core of this thesis work is a rather computationally efficient approach that consists of three processing stages. In the first, a set of (null) steering beamformers is used to carry out diverse partial signal separations, by using multiple arbitrarily located linear microphone arrays, each of them composed of a small number of microphones. In the second stage, each of the beamformer output goes through a classification step, which uses models for all the targeted sound classes (HMM-GMM, in the experiments). Then, in a third stage, the classifier scores, either being intra- or inter-array, are combined using a probabilistic criterion (like MAP) or a machine learning fusion technique (fuzzy integral (FI), in the experiments). The above-mentioned processing scheme is applied in this thesis to a set of complexity-increasing problems, which are defined by the assumptions made regarding identities (plus time endpoints) and/or positions of sounds. In fact, the thesis report starts with the problem of unambiguously mapping the identities to the positions, continues with AED (positions assumed) and ASL (identities assumed), and ends with the integration of AED and ASL in a single system, which does not need any assumption about identities or positions. The evaluation experiments are carried out in a meeting-room scenario, where two sources are temporally overlapped; one of them is always speech and the other is an acoustic event from a pre-defined set. Two different databases are used, one that is produced by merging signals actually recorded in the UPC¿s department smart-room, and the other consists of overlapping sound signals directly recorded in the same room and in a rather spontaneous way. From the experimental results with a single array, it can be observed that the proposed detection system performs better than either the model based system or a blind source separation based system. Moreover, the product rule based combination and the FI based fusion of the scores resulting from the multiple arrays improve the accuracies further. On the other hand, the posterior position assignment is performed with a very small error rate. Regarding ASL and assuming an accurate AED system output, the 1-source localization performance of the proposed system is slightly better than that of the widely-used SRP-PHAT system, working in an event-based mode, and it even performs significantly better than the latter one in the more complex 2-source scenario. Finally, though the joint system suffers from a slight degradation in terms of classification accuracy with respect to the case where the source positions are known, it shows the advantage of carrying out the two tasks, recognition and localization, with a single system, and it allows the inclusion of information about the prior probabilities of the source positions. It is worth noticing also that, although the acoustic scenario used for experimentation is rather limited, the approach and its formalism were developed for a general case, where the number and identities of sources are not constrained.

show abstract

Section: Missing Feature Techniquesmentioning

confidence: 99%