Goodness of pronunciation (GoP) is typically formulated with Gaussian mixture model-hidden Markov model (GMM-HMM) based acoustic models considering HMM state transition probabilities (STPs) and GMM likelihoods of context dependent phonemes. On the other hand, deep neural network (DNN)-HMM based acoustic models employed sub-phonemic (senone) posteriors instead of GMM likelihoods along with STPs. However, each senone is shared across many states; thus, there is no one-to-one correspondence between them. In order to circumvent this, most of the existing works have proposed modifications to the GoP formulation considering only posteriors neglecting the STPs. In this work, we derive a formulation for the GoP and it results in the formulation involving both senone posteriors and STPs. Further, we illustrate the steps to implement the proposed GoP formulation in Kaldi, a state-of-the-art automatic speech recognition toolkit. Experiments are conducted on English data collected from Indian speakers using acoustic models trained with native English data from LibriSpeech and Fisher-English corpora. The highest improvement in the correlation coefficient between the scores from the formulations and the expert ratings is found to be 14.89% (relative) better with the proposed approach compared to the best of the existing formulations that don't include STPs.

No abstract

No abstract

Automatic syllable stress detection is typically performed with a supervised classifier considering manually annotated stress markings and features computed within the syllable segments derived from phoneme transcriptions and their time-aligned boundaries. However, the manual annotation is tedious and the errors in estimating segmental information could degrade stress detection accuracy. In order to circumvent these, we propose to estimate stress markings in automatic speech recognition (ASR) framework involving finite-state-transducer (FST) without using annotated stress markings and segmental information. For this, we train the ASR system with native English data along with pronunciation lexicon containing canonical stress markings and decode non-native utterances as pronunciations embedded with stress markings. In the decoding, we use an FST encoded with the pronunciations derived using phoneme transcriptions and the instructions involved in a typical manual annotation. Experiments are conducted on polysyllabic words taken from ISLE corpus containing utterances spoken by Italian and German speakers and using the ASR models trained with three corpora. Among all the three models, the highest stress detection accuracies with the proposed approach respectively on Italian & German speakers are found to be 2.07% & 1.19% higher than and comparable to those with the two supervised classification approaches used as baselines.