📄 Building a Speech Corpus based on Arabic Podcasts for Language and Dialect Identification

Summary

Spoken language and dialect identification for Arabic is held back by a shortage of suitable speech data. The authors address this by assembling ArPod (Arpod-1.0), a corpus of a little over 8 hours of conversational audio taken from an Arabic podcast website. It covers MSA, English, and Saudi, Egyptian, Lebanese and Syrian dialects. Some of the recordings mix dialect with English. They split the material into a clean subset and a code-switched subset, then cut the test material into 6-, 30- and 60-second pieces. On this data they compared several classifiers: SVM, MLP, kNN and Extratrees on two hand-crafted acoustic/spectral feature sets, and a CNN that treats spectrograms as images. SVM with the larger Librosa-based feature set did best on both subsets, reaching F1 of about 96% on clean data and 98% on code-switched data with the shortest pieces. The spectrogram CNN lagged far behind. The resource is useful because it gives Arabic dialect research a freely shared, multi-speaker audio benchmark.

Key points

• Corpus contents: Arpod-1.0 totals 8:10 hours, recorded at 44.1 kHz with 16-bit encoding. Each language or dialect has between about 50 minutes and 1.5 hours, with two or more male and female speakers. The long files were cut into roughly 5-minute chunks with MKVToolNix, and the corpus is published on Kaggle.
• Two subsets:
  • Clean (3:40 h): MSA, ENG, KSA and SYR.
  • Code-switched (4:30 h): EGY, LEB and KSA-E, where speakers alternate between their dialect and English.
• Feature schemes:
  • Scheme 1 has 34 features, including 13 MFCCs, energy, zero crossing rate, spectral centroid, spread, entropy, rolloff and flux, and a 12-dimensional chroma vector.
  • Scheme 2 has 193 Librosa features: 40 MFCCs, a 128-band mel spectrogram, chroma, spectral contrast and tonnetz.
• Best results: SVM with scheme 2 scored F1 ≈ 96% on clean 6-second segments and ≈ 98% on code-switched 6-second segments. The spectrogram CNN reached only about 56% F1 on 1-minute clean segments.
• Effect of segment length:
  • On clean data, the feature-based models (kNN, SVM, Extratrees) got worse as segments grew longer. MLP was the exception and improved slightly.
  • On code-switched data, segment length had no visible effect on performance.

Interesting ideas

• I1. Sweeping one controlled variable (utterance length) across every model exposes trends that a single headline score hides, such as some models degrading on longer inputs.
• I2. Keeping "clean" and "mixed" (code-switched) data as separate test conditions, rather than pooling them, shows whether a model is robust to mixing or only to the easy case.
• I3. Running cheap classical classifiers alongside a deep model on the same data showed the simple SVM winning by a wide margin, which argues for always keeping strong lightweight baselines.
• I4. Comparing a compact feature set with a richer one on the same learner isolates how much the representation matters, independently of the model.
• I5. Splitting long raw recordings into uniform chunks before training makes uneven sources comparable and easier to balance.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (learn): Report parser and reader scores broken down by sentence-length buckets, so length-dependent failures are visible instead of averaged away.
• I2 (tree): Evaluate gparse and the perceptron parser separately on standard text and on informal or dialect-converted UD text, to see where conversion or register shifts hurt.
• I3 (learn): Keep the perceptron parser as a standing baseline next to gparse in every evaluation, so the transformer's gains must be demonstrated rather than assumed.
• I4 (tokens): Run ablations on the tagger and morphology with small and large explainable feature templates, to measure which features earn their cost.
• I5 (world): Cut long stories into fixed-size passages when building world-of-text test sets, so entity and event tracking can be compared across texts of very different lengths.