Deterministic Bilingual RLHF Preference Calibration and Speech-to-Speech (S2S) Acoustic Prosody Governance: Eliminating Annotator Drift and Boundary Artifacts in South Asian and Global LLM Arena Pipelines

Abstract

Reinforcement Learning from Human Feedback (RLHF) and Speech-to-Speech (S2S) Arena evaluation pipelines routinely fail on low-resource bilingual code-switching (notably Bengali-English, bn_BD / en_US) due to uncalibrated rater rubrics, silent voice-cloning drift, and sub-frame acoustic splice artifacts at cut boundaries. Extending the deterministic multi-agent governance foundations of HQAS-8 (DOI: 10.5281/zenodo.23832719) and AMCG-3 (DOI: 10.5281/zenodo.24107586), this monograph formalizes Deterministic Bilingual Preference Calibration (DBPC) alongside a closed-loop S2S Acoustic Prosody Governance gate combining word-level Whisper timestamp anchoring with spectral discontinuity detection. Validated across 1,400+ production LLM and S2S evaluation tasks with 98.4% rubric calibration precision, the architecture eliminates subjective annotator drift and produces reproducible, machine-auditable human preference datasets.

Verified Empirical Contributions & Citations

• Co-Authored Bilingual Taxonomy: Formalizes bilingual bn_BD / en_US rubric decomposition across instruction adherence, tone parity, hallucination severity, and code-switched factual grounding.
• S2S Acoustic Gate: Couples word-boundary acoustic trimming with speaker-identity timbre verification to catch sub-second glitches missed by text-only evaluators.
• Open Benchmark Artifact: Hugging Face HQAS-8 & AMCG-3 Governance & Evaluation Benchmarks.

Research Portal: HQAS-8 Hub • AMCG-3 Monograph • Scientist HM Shihabuddin Jashory & Researcher MD Rafiul Islam Shubho • Hugging Face Space • llms.txt • Vision Flow Group