📄 ASMDD: Arabic Speech Mispronunciation Detection Dataset

Summary

The paper addresses a gap in Arabic speech resources: there is little labelled data for spotting pronunciation errors, especially in children's speech. The authors release ASMDD, a corpus of isolated Arabic words spoken by 100 Egyptian children between 2 and 8 years old. The word list is the 100 words the authors identify as most frequent in Egyptian dialogue. Kids in nursery schools were shown each word and asked to say it. The recordings were made and cleaned of classroom noise in Audacity, then cut into one file per word. Expert listeners marked each word as correctly or wrongly pronounced, judging segmental errors. The paper trains no model and reports no detection scores. Its contribution is the resource itself, plus a per-speaker count of errors. The authors present it as raw material for fine-tuning self-supervised speech models such as wav2vec and HuBERT. It should also support tools that detect Arabic pronunciation mistakes, an area where child-speech data is scarce.

Key points

• 100 child speakers from Egypt, aged 2–8. Folders 00–30 hold all 100 words per child, and folders 31–99 hold only 50 words per child.
• Audio is mono, 44.1 kHz, 32-bit, recorded and denoised with Audacity (version 3.1.0 cited).
• Each word file carries a word ID (0–99) and a binary label. A mispronounced item gets "_N" appended to its index, and the folder name encodes gender and word count.
• Error counts per speaker vary widely among the full-list children, from 5 errors (speaker 27) to 36 (speaker 28), with speaker 13 at 29.
• The data is distributed via Mendeley Data (doi:10.17632/x54dg53rmr.1) and Google Drive.

Interesting ideas

• I1. Putting the gold label directly into the file name ("_N" for wrong) makes the dataset self-describing, so it can be split or counted without a separate annotation file. — For Mova (learn): silver-data files could carry a short suffix that records whether the witness parsers agreed or disagreed, so filtering stays trivial and transparent.
• I2. The authors fix a small, closed list of the most frequent items and collect it from many different producers, so errors become comparable item by item and person by person. — For Mova (tokens): a fixed probe set of very frequent English words and forms could be run across many text sources, showing which items the tagger and lemmatizer get wrong most often.
• I3. The paper reports error counts per speaker rather than one overall rate, which exposes the strong spread between individuals. — For Mova (learn): evaluation could report errors per document or per source genre, so outlier texts that drag accuracy down become visible instead of hidden in an average.
• I4. The corpus deliberately keeps real erroneous productions alongside correct ones, under a cheap binary right/wrong label. — For Mova (answer): graded readers could be tested on a set of genuinely wrong student-style answers marked only right/wrong, which would act as negative controls for the checking logic.