📄 A Novel Speech Analysis and Correction Tool for Arabic-Speaking Children

Summary

The authors present ArPA (Arabic Pronunciation Aid), an application for Arabic-speaking children who struggle to pronounce letters of the alphabet. They argue that earlier mispronunciation detectors target adults or non-native learners, or cover only narrow error types such as short vowels. ArPA has two parts. A diagnostic module cleans the recording with Gaussian-filter denoising and silence removal, extracts MFCC and Mel-spectrogram features, renders them as colour images, and classifies each utterance as correct or mispronounced. A therapeutic module is a voice-only game in which an avatar advances a level for each correct letter, and progress reports are available to parents and therapists. Two datasets were used: one recorded at a childcare centre and one of Arabic alphabet pronunciations. Classical baselines (KNN, SVM, Decision Trees) on MFCC features were compared with a fine-tuned ResNet18 on image-converted features. ResNet18 on Mel-spectrogram images performed best, reaching 99.015% accuracy on Dataset 2. The work matters as a child-focused tool that couples automatic diagnosis with motivating practice, although the authors acknowledge that their data is small.

Key points

• Pipeline: denoising with a Gaussian filter, then silence removal, then MFCC or Mel-spectrogram extraction (pre-emphasis, 20–40 ms Hamming-windowed frames, FFT, Mel filter bank, log, plus DCT for MFCC), then min–max colormap conversion to images, then a binary correct/incorrect decision.
• Classical baselines on MFCC features: KNN was the strongest on both datasets. On Dataset 2 the three baselines spanned roughly 82%–94% accuracy, and SVM struggled with the small data.
• ResNet18 on MFCC images reached 97.53% accuracy on Dataset 1 and 96.24% on Dataset 2.
• ResNet18 on Mel-spectrogram images reached 98.77% accuracy on Dataset 1 and 99.015% on Dataset 2.
• Training setup:
  • random 80/20 train/test split;
  • minibatch size 100, 30 epochs;
  • learning rate 0.0001 with the Adam optimizer;
  • results reported as precision, recall, F1, accuracy and per-class confusion matrices.

Interesting ideas

• I1. The uncompressed Mel-spectrogram beat MFCC, its DCT-compressed derivative, and the authors attribute this to the extra time-frequency detail the classifier could exploit. — For Mova (tokens): when building tagger or morphology features, test whether keeping a fuller representation (e.g. all morphological feature values rather than a collapsed tag) improves accuracy before settling on compact encodings.
• I2. Recasting a numeric feature matrix as an image, through a simple min–max colormap mapping, lets an off-the-shelf pretrained vision backbone be reused instead of designing a new architecture. — For Mova (tree): gparse experiments could encode pairwise head-dependent score matrices in a form that existing pretrained encoders accept, as a cheap way to borrow capacity.
• I3. Cheap classical models (KNN, SVM, Decision Trees) were run first as a baseline floor, so the gain from the deep model is measurable rather than assumed. — For Mova (learn): every new learned component should report a simple baseline alongside it, which acts as a sanity control when judging improvements from silver data or LLM teachers.
• I4. Diagnosis drives practice directly: each letter the child fails gets targeted exercises, and per-item progress is logged for a supervising adult. — For Mova (answer): student readers could keep a per-skill mastery log across school-course tests and generate focused drills on the question types they get wrong, with an explainable report.
• I5. Evaluation goes beyond overall accuracy to per-class precision, recall and confusion matrices, which confirms that both the "correct" and "mispronounced" classes are handled well. — For Mova (learn): report per-label confusion for dependency relations or semantic roles so that strong majority classes cannot hide failures on rare ones.