Jun 5, 2026

Three Hugging Face datasets from CMU's TartanAviation air traffic control corpus: paired audio and ADS-B clips, a half-million-utterance VAD split, and a full ASR label set. Why decode-time biasing and LLM correction lost to a plain ROVER vote, and the one signal that moved label quality.

Jun 4, 2026

Benchmarking the Gemma 4 family (E2B, E4B, and the new 12B) across all eight Open ASR Leaderboard short-form datasets, batched. WER, RTFx, where a general-purpose multimodal LLM holds up, and where it doesn't.