National Technical University of Athens (NTUA)
Μέθοδοι εύρωστης αναγνώρισης ομιλίας με πολλαπλούς αισθητήρες και εφαρμογές σε ευφυή περιβάλλοντα
Abstract
dc:descriptionAutomatic speech recognition is extensively used in human-computer interaction mainly because communication via speech is natural, fast, and safe. Speech applications and voice-enabled, hands-free technologies are increasingly integrated to the modern area of multisensory smart environments (e.g., smart homes, assistive robotics), targeting life wellness and effectiveness at work. The recent advancements of deep learning have brought significant improvements, yielding word error rates below 10% in most cases of recognition with close-talk microphones. However, many challenges remain unsolved, especially in Distant Speech Recognition (DSR) applications where noise and reverberation effects degrade significantly the performance.In this dissertation, we examine a combination of multichannel methods for robust DSR using microphone arrays. More specifically, the proposed methodology includes 1) microphone selection using signal-based criteria of confidence, 2) microphone combination in the decision level following a hypothesis rescoring approach, 3) data augmentation and multi-style training by simulating domestic acoustic scenes with noisy far-field speech for applications in environments with limited amount of training data, 4) multi-microphone adaptation schemes, and 5) extraction of non-linear AM-FM features that are combined with traditional energy-based features. A multichannel demodulation approach is proposed for the extraction of improved frequency modulation features that capture the fine structure of speech formants and constitute beneficial and supplementary to the traditional energy-based cepstral features (e.g, MFCCs). Improvements have been demonstrated mainly in GMM-HMM systems for small and large vocabulary tasks. Yet, they have limited applications in DNN-HMM systems and DSR tasks. Herein, we elaborate on their integration within state-of-the-art front-end schemes that include tansformation of MFCCs resulting in discriminant and speaker-adapted features of large temporal contexts. We explore 1) multichannel demodulation schemes for multi-microphone setups, 2) richer descriptors of frequency modulations, and 3) feature transformation and combination via hierarchical deep bottleneck networks. We present results for tandem and hybrid recognition with GMM and DNN acoustic models, respectively. The improved modulation features are combined efficiently with MFCCs yielding modest and consistent improvements in widely known and challenging corpora for multi-microphone DSR benchmarking.Based on the proposed multichannel methods, a practical system pipeline is designed for always-listening, far-field spoken command recognition in every-day smart indoor environments of multiple rooms equipped with sparsely distributed microphone arrays. The system consists of a traditional cascade of far-field spoken command detection and recognition, the former based on the detection of activating key-phrases. A room-dependent speech activity detection module is used to enable parallel recognition pipelines in each room. The proposed approach is developed for the Greek language, exhibiting promising performance in real and challenging domestic acoustic scenes, reaching 76.6% command recognition accuracy, a result that represents 46% relative improvement over conventional beamforming.An online version of the proposed spoken command recognition system is implemented for four languages (Greek, German, Italian, English) and combined with gesture recognition, aiming at multimodal human-robot interaction. Within the context of assistive robotics, we develop a smart and natural interface that provides multimodal sensory processing capabilities. By fusing the complementary unimodal information streams, we obtain the optimum multimodal hypothesis. Additionally, we explore new aspects on assistive living via smart social human-robot interaction (HRI). We provide a framework of resources, including datasets and tools that are demonstrated in two real-life use cases for elderly subjects: a multi-modal interface of an assistive robotic rollator and an assistive bathing robot, where following the proposed fusion scheme, the obtained command accuracy reaches 90% in real interaction scenarios with elderly users.
Degree
thesis:*- Grantor dc:publisher
- National Technical University of Athens (NTUA)
- Year dc:date
- 2018
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Rodomagoulakis, Isidoros
- Ροδομαγουλάκης, Ισίδωρος
Subjects
dc:subject × 10- Αυτόματη αναγνώριση ομιλίας
- Επικοινωνία ανθρώπου - μηχανής
- Ευφυή περιβάλλοντα
- Automatic speech recognition (ASR)
- Human robot interaction
- Smart environments
- Επιστήμες Μηχανικού και Τεχνολογία
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Engineering and Technology
- Electrical Engineering, Electronic Engineering, Information Engineering
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/44766