i-Vector Modeling of Speech Attributes for Automatic Foreign Accent Recognition

Abstract

We propose a unified approach to automatic foreign accent recognition. It takes advantage of recent technology advances in both linguistics and acoustics based modeling techniques in automatic speech recognition (ASR) while overcoming the issue of a lack of a large set of transcribed data often required in designing state-of-the-art ASR systems. The key idea lies in defining a common set of fundamental units "universally" across all spoken accents such that any given spoken utterance can be transcribed with this set of "accentuniversal" units. In this study, we adopt a set of units describing manner and place of articulation as speech attributes. These units exist in most spoken languages and they can be reliably modeled and extracted to represent foreign accent cues. We also propose an i-vector representation strategy to model the feature streams formed by concatenating these units. Testing on both the Finnish national foreign language certificate (FSD) corpus and the English NIST 2008 SRE corpus, the experimental results with the proposed approach demonstrate a significant system performance improvement with p-value < 0.05 over those with the conventional spectrum-based techniques. We observed up to a 15% relative error reduction over the already very strong i-vector accented recognition system when only manner information is used. Additional improvement is obtained by adding place of articulation clues along with context information. Furthermore, diagnostic information provided by the proposed approach can be useful to the designers to further enhance the system performance.

A Machine Hearing System for Robust Cough Detection Based on a High-Level Representation of Band-Specific Audio Features

“…In this regard, the use of a simple approach based solely on mean and standard deviation of the short time features has shown good performance compared to the more complex BoAW. More complex approaches using contextual information such as ivectors [39] could also be explored. However, the particular context (continuous, smartphone-based monitoring with low battery consumption) would not benefit from this approach.…”

Hybrid Method for Stress Prediction Applied to GLAFF-IT, a Large-Scale Italian Lexicon

“…current version of GLAFF-IT counts 37, 320 lemmas for 457, 702 wordforms and includes nouns, verbs, adjectives and adverbs. 2 Each entry of the lexicon includes a wordform, a tag in MULTEXT-GRACE format [13] specifying the main syntactic category and inflection features, a lemma and API phonological transcriptions with the stress placement when present in GLAW-IT. An extract of GLAFF-IT is reported in Figure 2.…”

“…We use a model that has learned the phonological contexts for stressed and unstressed Italian vowels. Orthographic and phonological context-based approaches have been extensively used in the text-to-speech domain for stress detection [2,7] and for accenting unknown words in a specialised language [18]. The rationale behind our approach is that the exploitation of the phonological neighbourhood of a vowel helps estimate its probability of being stressed or unstressed.…”

Automatic Characterisation of the Pronunciation of Non-native English Speakers using Phone Distance Features

“…Approaches employed so far in the literature for L1 classification include i-vector modelling [14], GMMs trained on MFCCs [15], and prosodic [16] approaches, with varying degrees of success. The approach investigated in this paper predicts, from recordings of spontaneous speech, the speaker's native language (L1) from among 21 different languages and, in the case of Spanish speakers, their country of origin from among three countries.…”