Lower WER Research Articles

Arabic Question Generation Using Transformers

After the increased reliance on online education, online assessment became an essential tool for educators to remotely monitor and evaluate students’ understanding in order to assist them properly. However, the laborious process of creating exam questions is a challenge for most teachers. Thus, automated Question Generation aims to assist teachers by generating questions from given data. Limited research has been conducted to tackle this issue in the Arabic Language due to the complexity of the language and the limited amount of available Arabic data. This paper explores different implementations of the transformer models, that demonstrated their superiority in natural language processing. Three approaches were introduced to tackle this problem with Arabic data using an Arabic-based transformer, an English-based transformer, and a multilingual-based transformer. Each of the fine-tuned models was trained using ARCD, XGLUE, DialectBench, and ArabicQA data sets and evaluated on automatic and manual metrics. Two of the proposed models achieve state-of-the-art results on the Arabic question generation task. The English transformer obtained a ROUGE score of 0.59 on XGLUE, while the Arabic transformer model achieves 0.49 on ARCD. Both of these models demonstrate excellent quality of questions through human-conducted evaluations by achieving low WER and high GC, U, and A scores.

PaSCoNT - Parallel Speech Corpus of Northern-central Thai for automatic speech recognition

This paper proposed a Parallel Speech Corpus of Northern-central Thai (PaSCoNT). The purpose of this research is not only to understand the different linguistic characteristics between Northern and Central Thai, but also to utilize this corpus for automatic speech recognition. The corpus is composed of speech data from dialogues of daily life among northern Thai people. We designed 2,000 Northern Thai sentences covering all phonemes, in collaboration with linguists specialized in the Northern Thai dialect. The samples in this study are 200 Northern Thai dialect speakers who had been living in Chiang Mai province for more than 18 years. The speech was recorded in both open and closed environments. In the speech recording, each speaker must read 100 pairs of Northern-Central Thai sentences to ensure that the speech data comes from the same speaker. In total, 100 h of speech were recorded: 50 h of Northern Thai and 50 h of Central Thai. Overall, PaSCoNT consists of 907,832 words and 6,279 vocabulary items. Statistical analysis of the PaSCoNT corpus revealed that 49.64 % of words in the lexicon belongs to the Northern Thai dialect, 50.36 % from the Central Thai dialect, and 1,621 vocabulary items appeared in both Northern and Central Thai. Statistical analysis is used to examine the difference in speech tempo, i.e. time per phoneme (TTP), syllable per minute (SPM), between Northern and Central Thai. The results revealed that there were statistically significant differences speech tempo between Central and Northern Thai. The TTP speaking and articulation rate of Central Thai is lower than Northern Thai whereas SPM speaking and articulation rate of Central Thai is higher than Northern Thai. The results also showed that the ASR model training using Northern Thai speech corpus provides the lower WER% when testing using Northern Thai testing speech data and provides the higher WER% when testing using Central Thai Testing speech data and vice versa. However, the ASR model training using the PaSCoNT speech corpus provides the lower WER% for both Northern Thai and Central Thai testing speech data.

Heterophonic speech recognition using composite phones.

Heterophones pose challenges during training of automatic speech recognition (ASR) systems because they involve ambiguity in the pronunciation of an orthographic representation of a word. Heterophones are words that have the same spelling but different pronunciations. This paper addresses the problem of heterophonic languages by developing the concept of a Composite Phoneme (CP) as a basic pronunciation unit for speech recognition. A CP is a set of alternative sequences of phonemes. CP’s are developed specifically in the context of Arabic by defining phonetic units that are consonant centric and absorb phonemically contrastive short vowels and gemination, not represented in the Arabic Modern Orthography (MO). CPs alleviate the need to diacritize MO into Classical Orthography (CO), to represent short vowels and stress, before generating pronunciation in terms of Simple Phonemes (SP). We develop algorithms to generate CP pronunciation from MO, and SP pronunciation from CO to map a word into a single pronunciation. We investigate the performance of CP, SP, UG (Undiacritized Grapheme), and DG (Diacritized Grapheme) ASRs. The experimental results suggest that UG and DG are inferior to SP and CP. For the A-SpeechDB corpus with MO vocabulary of 8000, the WER for bigram and context dependent phone are: 11.78, 12.64, and 13.59 % for CP, SP_M (SP from manual diacritized CO), and SP_A (SP from automated diacritized MO) respectively. For vocabulary of 24,000 MO words, the corresponding WER’s are 13.69, 15.08, and 16.86 %. For uniform statistical model, SP has a lower WER than CP. For context independent phone (CI), CP has lower WER than SP.

Automatic multimedia indexing: combining audio, speech, and visual information to index broadcast news

This paper describes an indexing system that automatically creates metadata for multimedia broadcast news content by integrating audio, speech, and visual information. The automatic multimedia content indexing system includes acoustic segmentation (AS), automatic speech recognition (ASR), topic segmentation (TS), and video indexing features. The new spectral-based features and smoothing method in the AS module improved the speech detection performance from the audio stream of the input news content. In the speech recognition module, automatic selection of acoustic models achieved both a low WER, as with parallel recognition using multiple acoustic models, and fast recognition, as with the single acoustic model. The TS method using word concept vectors achieved more accurate results than the conventional method using local word frequency vectors. The information integration module provides the functionality of integrating results from the AS module, TS module, and SC module. The story boundary detection accuracy was improved by combining it with the AS results and the SC results compared to the sole TS results

Depressive symptomatology during pregnancy and the postpartum period: Is the beck depression inventory applicable?

Abstract The Beck Depression Inventory (BDI), a widely used tool to assess depression during pregnancy and the postpartum period, has some item heterogeneity and contains some items that may be normative during pregnancy rather than uniquely indicative of depression. To evaluate the implications of this, we studied 62 primiparous women and a comparison group of 36 nulliparous women. Assessments were made at the 31st week of pregnancy and 12 weeks postpartum in the pregnant group, as well as at similar intervals in the nonpregnant group. The BDIitems were divided into two conceptually based clusters: Cognitive-Affective and Somatic. A third cluster was designated Residual. This conceptual grouping was supported by moderately high alpha coefficients and low intercorrelations between clusters. The total BDI and the clusters were then evaluated longitudinally in two analyses: (1) pregnant and nonpregnant women with low psychiatric risk were compared, and (2) pregnant women of low and high psychiatric risk wer...