Audio-Laya: direct speech-to-decision classification with a frozen Whisper encoder and a trainable projector

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Research prototype that classifies call audio as legitimate or spam directly from audio, without an intermediate transcription step. A frozen whisper-small.en encoder feeds a trainable 768-to-1024 projector whose output is appended to frozen ModernBERT text states inside the pretrained Laya decision model, so only the projector (and optionally Laya's decision components) is trained. The deposit contains the source code, configuration, tests, the stored evaluation results, and the technical reports in English and Thai. It reports the internal call-level split, cross-task zero-shot transfer, semantic-alignment retrieval, and an external false-alarm check, together with the limitations that restrict what those numbers support. Headline results. On a 282-call internal test split, projector-only adaptation reached 99.6% macro F1 and projector-plus-decision-head reached 100.0%, against 89.2% for a Whisper-ASR-to-Laya text baseline. Zero-shot transfer to MInDS-14 14-way intent classification stayed at or below the majority-class baseline for both direct-audio variants, while the transcript route reached 93.6%. Limitations. All positive examples come from the FTC robocall corpus and all negative examples from HarperValleyBank, so corpus, recording conditions, and label are completely confounded. The high in-domain scores are a pipeline comparison, not evidence of real-world spam detection, and a matched-source, source-disjoint evaluation is required before any such claim. The model is not calibrated, and the urgency probe is a 40-clip synthetic set whose three-epoch result was selected after an earlier evaluation on the same clips. Not included. No source audio is redistributed; the datasets are linked as related identifiers and the README documents how to rebuild the manifests. Trained checkpoints are published separately on Hugging Face. The technical reports are released under CC BY 4.0; the code, configuration, and scripts are released under the MIT License.