Manuscript hardening
Counterfactual Decision-Stability ASR
A speech-to-decision stability research line that asks whether downstream decisions remain stable under plausible transcript alternatives.
Thesis
ASR quality should be evaluated not only by word error, but by whether plausible recognition alternatives change downstream decisions in high-stakes conversational workflows.
Why it matters now
Real deployments use transcripts as input to decisions, summaries, and escalation logic. A transcript can look acceptable while still changing the action a downstream system takes.
Evidence surface
- JANUS workspace organized around audio, ASR hypotheses, risk atoms, counterfactual variants, CEIS scoring, and constrained recovery.
- Selected-300 aggregate review surface completed with dual-review and model-level assessment summaries.
- Evidence-chain audits preserve aggregate counts and readiness gates while keeping audio, transcripts, row IDs, and reviewer notes outside the public site.
Validation path
- ASR hypotheses
- Risk atoms
- Counterfactual variants
- Decision stability
Current outputs
Manuscript evidence, aggregate audits, recovery-policy checks, and submission-readiness artifacts.
Scope control
Public pages describe the research question and aggregate readiness only. Raw audio, transcripts, row content, local hypotheses, and reviewer notes remain private.
Questions
- Which transcript alternatives are plausible enough to matter for downstream review?
- When does low WER still hide decision-relevant risk?
- What recovery policy is conservative enough for high-stakes speech workflows?