📄 SPARTA: Speaker Profiling for ARabic TAlk

Summary

The paper addresses speaker profiling from Arabic audio. Its goal is to infer three traits of the person talking: gender, emotional tone and dialect. Arabic speech resources are scarce, so the authors pooled six publicly obtainable corpora. They cleaned and re-segmented the data and published fixed train/dev/test splits as an open benchmark. They then compared two kinds of input. The first is raw acoustic features (MFCC, MEL) fed to CNN and LSTM models. The second is pretrained utterance embeddings (i-vectors, d-vectors, x-vectors and their concatenations) fed to a fully connected network. Each setup was trained either per task (STL) or jointly with a shared trunk and per-task outputs (MTL). The best system concatenates all three embedding types and uses MTL. Joint training helped most for emotion and dialect, while gender was close to saturated. Compact pretrained representations clearly beat raw features in this low-data regime. On KSUEmotion, the system exceeds previously reported emotion and gender results. The released splits and models give Arabic speech classification a reproducible baseline.

Key points

• Data: six source corpora (KSUEmotion, ANAD, QCRI dialect data, SARA, MDAS, KING), each labelled for only a subset of the three traits. The splits are released at github.com/mawdoo3/sparta-benchmark.
• Label sets:
  • Dialect uses five classes: EGY, NOR, GLF/GUL, LEV/LAV, MSA.
  • Emotion follows the KSUEmotion scheme: neutral, happy, sad, angry, surprised, questioning.
• Best configuration: a fully connected network on idx-vector (i+d+x concatenation) with MTL. Its overall F1 is 76.01, against 73.89 for STL. Per task, MTL reaches emotion F1 70.16 (STL 66.53) and dialect F1 58.44 (STL 56.05). Gender peaks at 99.81 (dx-vector, MTL).
• Raw-feature models lag behind: CNN/LSTM on MFCC or MEL reach only about 21–29 F1 on dialect. The authors attribute this to data hunger and to weakness on long sequences.
• Comparison with prior work: earlier work on KSUEmotion reported 79.3% accuracy for emotion and 98.7% for gender. The proposed system beats both.

Interesting ideas

• I1. Several corpora, each annotated for only some of the target traits, are merged into one multi-task model with a shared trunk and per-task heads, so every corpus contributes to the shared representation.
• I2. The weakest embedding on its own (d-vector) still raised scores once concatenated with the others, because it was trained on different data and holds complementary information.
• I3. Fixed public splits are published for every constituent dataset, and scores are reported per dataset as well as for the pooled set, which exposes where gains actually come from.
• I4. With little data, compact pretrained representations far outperformed models that learn from raw low-level input. The gap was largest on the harder tasks (emotion, dialect).
• I5. The emotion taxonomy places "questioning" next to affective states, effectively labelling speech act and affect in one scheme.

For Mova

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

• I1 (learn): train tagger, parser and role heads together on treebanks and silver sets that carry only partial annotation layers, so each layer's data strengthens the shared encoder.
• I2 (tree): when combining encoder features in gparse, score each source both alone and inside the combination before dropping it, since a weak solo feature may still add unique signal.
• I3 (learn): report parser and reader metrics per treebank or test source, not only pooled, so that a gain on one domain cannot hide a regression on another.
• I4 (tokens): for rare morphology or tagging classes with few examples, prefer features from pretrained components over training from raw characters, and measure the difference explicitly.
• I5 (frames): predict speech act and affect/irony as sibling heads over the same utterance in prag, letting cues for one inform the other while keeping both outputs explainable.