Audio Flamingo: A Novel Audio Language Model with Few-Shot Learning and Dialogue Abilities

Existing audio LMs could not combine broad audio understanding with in-context learning and multi-turn dialogue.

Builds on a frozen/fine-tuned Flamingo-style Perceiver-resampler architecture bridging audio encoders to an LLM, plus training techniques and data strategies targeting in-context learning and multi-turn dialogue. Audio is again compressed to a fixed token budget by the resampler.

ICML 2024 per the arXiv comment. The successor line (Audio Flamingo 3, 2507.08128; Audio Flamingo Next, 2604.10905) extends this work; none of these are in production serving stacks. Consistent with the speech-LM cluster: no paper in this family reports a hardware-level latency or throughput measurement, which is itself a signal that the speech-LM literature is optimising for capability demos rather than serving economics. VENUE_TRACK BACKFILL 2026-10-04 (paper-fields pass, agent w3p): venue_track 'main'. This is STRUCTURAL rather than researched: icml has no Findings volume and no workshop track in which a paper could appear, so 'main' is the only possible honest value and no per-paper evidence was needed. The field is annotated here so that 'findings' and 'workshop' remain MEANINGFUL - a filter for findings=true must return only the papers that actually are findings, which requires every other paper to be explicitly NOT findings.