Building ASR + LLM + RAG pipelines and agent workflows for conversational analysis, structured extraction, and evidence-aware reasoning over long-form audio and transcripts.
Current thesis
Speech workflows become genuinely useful when they preserve grounding across audio, transcript, retrieval, and analyst judgment. The interesting work is not another chat wrapper around transcripts; it is a pipeline that keeps evidence aligned while moving toward decisions.
2026 perspective
The 2026 shift is away from short-clip demos and toward long-form, multi-speaker, evidence-carrying audio workflows. Speech systems now need to preserve what transcripts flatten.
Human factors lens
Real users do not want to babysit a black box over a 40-minute call. They need segment-level grounding, fast correction loops, and outputs that match the time pressure of analyst work.
LongSpeech, released on January 20, 2026, makes the gap explicit by showing how quickly model quality drops once tasks involve long audio, multi-task reasoning, and sustained context. In 2025, WavRAG argued that transcript-only RAG throws away useful audio information, CORAL showed that real conversational retrieval is multi-turn and citation-sensitive, and recent EMNLP work on context discovery for ASR demonstrated that retrieval can improve rare-term recognition with lower latency than heavier LLM-only alternatives. Taken together, these signals shift the research question from recognition alone to end-to-end workflow design.
Keep links from raw audio to transcript spans to retrieved passages to final claims.
Use transcript retrieval where it is sufficient, but keep audio-aware retrieval for cases where prosody, speaker state, or transcription error matters.
Let humans verify segments, entities, and extracted fields locally instead of reviewing a large final answer only at the end.
Generate summaries, entities, timelines, and alerts in formats that downstream teams can act on or audit.
Measure rare-term accuracy, speaker attribution, and robustness across noisy, accented, or domain-shifted speech.
Evaluate whether retrieved passages or audio segments actually support the generated claim, citation, or extraction.
Test performance across topic shifts, long meetings, and multi-turn conversations where memory and retrieval interact.
Track correction time, number of manual fixes, and how often humans must return to raw audio.
An arXiv benchmark released on January 20, 2026 that highlights the practical difficulty of transcription, translation, and reasoning over long speech.
ACL Anthology work from 2025 arguing that transcript-only RAG can discard useful acoustic evidence in spoken-dialogue systems.
A 2025 conversational RAG benchmark that makes multi-turn retrieval and attribution more realistic than single-turn QA setups.
EMNLP 2025 work showing retrieval-augmented context discovery can improve rare-word recognition while keeping latency practical.