Numerous mispronunciation detection and diagnosis (MD&D) research methods attempt to utilize both auditory and linguistic input elements. However, the absence of a substantial quantity of annotated training data at the phoneme level constrains the improvement of performance. Recent advancements in speech recognition, such as the LAS-Transformer, which is an enhanced Transformer based on the Local Attention Mechanism, suggest potential applications in improving mispronunciation detection by utilizing more advanced attention mechanisms in processing auditory inputs ^[1]. To construct a more robust MD&D system, authors combined the embedding properties of acoustic, phonetic, and linguistic data, abbreviated as APL. In ^[2], the suggested method has a detection accuracy of 9.93% higher than the baseline, a diagnosis error rate of 10.13% lower, and an F-measure of 6.17% higher than the baseline. These results were acquired through experimental work performed on the L2-ARCTIC database. The authors introduced a phoneme-level MD&D system that utilizes acoustic embedding (acoustic characteristics), phonetic embedding, and linguistic embedding (canonical phoneme sequence) as inputs in order to predict the spoken phoneme sequence. In ^[3], an acoustic-graphemic phonemic model (AGPM) utilizing multi-distribution deep neural networks (MD-DNNs) is proposed, whose input features consist of acoustic data, graphemes, and canonical transcriptions (encoded as binary vectors). The AGPM is capable of intuitively modeling both grapheme-to-likely-pronunciation and phoneme-to-likely-pronunciation conversions, which are incorporated into acoustic modeling. Using the AGPM, a unified MDD framework that functions similarly to freephone recognition is constructed. Experiments indicate that the proposed technique yields an 11.1% phone error rate (PER). The false rejection rate (FRR), the false acceptance rate (FAR), and the diagnostic error rate (DER) for MDD are 4.6%, 30.5%, and 13.5%, respectively. While this model showed promising results, it is notable that it resulted in relatively high false rejection and acceptance rates, indicating room for improvement in model accuracy.

Computer-aided pronunciation training systems necessitate reliable automated pronunciation error detection techniques to recognize human faults. Yet, the overall number of mispronounced speech data utilized to train these algorithms and their manual annotation reliability greatly affect their performance ^[4]. To resolve this issue, the authors in ^[4] employed anomaly detection methods to identify mispronunciation. Their anomaly detection model was the One-Class SVM, using phoneme-specific models. A bank of binary DNN speech attribute detectors retrieved manners and locations of articulation for each model. Multi-task learning and dropout were implemented to reduce DNN speech attribute detector overfitting. The model was trained using the WSJ0 and TIMIT standard datasets, which contain solely native English speech data, and then assessed it using three datasets: a native English speaker corpus with fake mistakes, a foreign-accented speech corpus, and a children’s disordered speech corpus. Lastly, the proposed approach was compared to the usual goodness-of-pronunciation (GOP) algorithm to prove its efficacy. The technique lowered false-acceptance and false-rejection rates by 26% and 39% compared to the GOP technique. Furthermore, in ^[5], the authors presented a speech recognition system capable of detecting mispronunciations. The dataset contains 89 students, 46 of whom are female. Ten times 28 Arabic phonemes are voiced. MFCCs are retrieved from 890 utterances for modeling with five different machine-learning models. K nearest neighbor (KNN), support vector machine (SVM), naive Bayes, multi-layer perceptron (MLP), and random forest (RF) are some of them. The experimental findings show that the random forest approach achieves an accuracy rate of 85.02%, which is higher than that of other machine learning models. Shareef et al. ^[6] emphasize the extensive comparison of feature extraction algorithms for the purpose of identifying impaired Arabic speech. The feature extraction approach is based on several wavelet transformation variants. LSTM and CNN-LSTM models are built to identify the impairment in Arabic speech. The combination of MFCCs and LSTM achieves the highest classification accuracy (93%), followed by CNN-LSTM (91%).

Table 2 summarizes the various approaches used in these studies for mispronunciation detection and diagnosis. However, these studies have largely overlooked the influence of gender on pronunciation, an aspect that may hold key insights for personalized language learning.