DASER-Net

Research investigating disentangled representations for speech emotion recognition across distinct corpora. My contribution: real-corpus data validation and speaker-disjoint evaluation.

Organisation: INFERENCE Lab (research collaboration · since Aug 2026)
Role: Collaborator: real-data validation & reproducible evaluation
Timeline / Repo: Aug 2026 – present

Traceroute path · 6 hops

Each hop isolates problem, constraints, architecture decisions, verification artifacts, and limits.

- If the same speaker or recording room appears in both train and test folds, validation metrics measure speaker recognition rather than emotion generalization.
- Real multi-corpus training requires deterministic manifest freezing before comparing any neural architecture.
Next hop follows below
- The model architecture and initial synthetic pilot were authored by lead researcher Muhammad Khubaib Ahmad; my remit is strictly real data integrity.
- Collaboration repository (Inference-LAB/DASER-Net) returned HTTP 404 to public unauthenticated requests on 11 Nov 2026 (private/restricted access).
Next hop follows below
- Done: built CREMA-D ingestion with integrity checks for duration, sample validity, and metadata consistency.
- Done: generated deterministic speaker-disjoint train/validation/test manifests so no speaker spans two partitions.
- Current: validating an audio baseline to sanity-check end-to-end feature loading prior to full cross-corpus runs.
Structural schematic · Illustrative (not measured telemetry)
Illustrative planned evaluation matrix (train on Corpus A -> evaluate unseen Corpus B). Diagram is structural, not benchmark output.
Next hop follows below
- Each pipeline stage carries a standalone validation script inside the collaboration repo.
- Note: no public GitHub link is clickable because the repository is currently private (verified HTTP 404 on 11 Nov 2026).
Asset status (screenshot): Collaboration repository (Inference-LAB/DASER-Net) is currently private; no synthetic results or unconfirmed scores are published.
Next hop follows below
- Only a synthetic pilot ran earlier in the repo; synthetic numbers do not demonstrate real speech generalization and are excluded.
- No publication or authorship claims are asserted at this stage.
Next hop follows below
- Verify dataloader determinism across seeds and record first reproducible real-corpus F1/UAR baselines once approved for disclosure.