What's Happening In Accents & Dialects ? A Review Of The State Of The Art (post-Interspeech 2013) Martin Russell – University of Birmingham Andrea DeMarco – University of East Anglia Christophe Veaux – University of Edinburgh Maryam Najafian – University of Birmingham Overview of Themes 1) Classification & Identification – Andrea 2) Speech Synthesis - Christophe 3) Automatic Speech Recognition – Maryam 4) Human Perception and Production – Maryam Classification & Identification Languages, accents & dialects ● A total of 11 papers surveyed (not a lot) ● Various application scenarios, but most work is on ● Language Identification (LID) We'll have a look at: ● – Feature extraction techniques – Classification methods – Corpora – Results – What's happening next Classification & Identification - Application Scenarios Foreign Accent Detection from Spoken Finnish [5] ● Native British Accent Classification [7] ● Accent Quantification of Indian Speakers of English ● [11] Language Identification [1,2,3,4,6,8,9,10] ● Classification & Identification - Feature Extraction MFCC → RASTA → CMVN → VTLN → SDC ● MFCC → Warping X ~ N(0,I) → SDC → Concatenate ● MFCC → Delta → Delta-Delta → CMVN ● Phone lattices and n-grams, absolute (what) and ● relative (where) distance kernels (PARF) Phone Log-Likelihood Ratios (PLLR) → PCA ● Phonotactic i-Vectors ● Classification & Identification - Classification Methods i-Vectors – a point estimate of an utterance in variability ● subspace Speaker Compensation ● – Linear/Semi-supervised/Heteroscedastic/Probabilistic Discriminant Analysis – Neighbourhood Component Analysis Binary Genetic Algorithm-based classifier fusions ● Traditional GMM models for supervised phoneme classes ● SVM Kernels ● DARPA RATS ANN on i-vectors ● – 3 layers, i-vector input, 6-language posterior output – 400-700 hidden nodes DARPA RATS Adaptive Gaussian Backend ● Classification & Identification - Corpora FSD (Finnish National Foreign Language Certificate ● Corpus) ABI (Accents of the British Isles Corpus) ● Custom Indian Speaker Dataset ● NIST Language Recognition Evaluation (LRE) ● RATS LID Data Corpus (5 targets, 10 non-targets) ● Classification & Identification - Results Corpus Novel Method Baseline FSD (iVector) 20.01% EER 24.13% EER ABI (iVector) 81% Accuracy 73.6% Accuracy LRE (PARF) 19.89% EER (3s test) 23.90% EER (3s test) LRE (PLLR) 3.21% C ,1.79% C 3.79% C ,2.09% C avg avg avg avg RATS (iVector-ANN) 6.95% EER 8.99% EER LRE (Phon. iVector) 19.11% EER (3s test) 22.60% EER (3s test) RATS (iVector-AGB) 3.6% C (30s test) 4.9% C (30s test) avg avg ●Indian accent strength (like in other languages) can be tied down to models of specific phonemes – mostly consonants in Indian. Machine performance equalled human listeners. Take Home Message (1) Feature Vector Overview for TRAP Language ● Identification System for RATS Phase II Evaluation Take Home Message (2) Different factor sizes/UBM components/Dim ● Reductions. Classifiers behave differently – Fusion gives a big boost (Accents of the British Isles)