bosonai/higgs-audio-v3-8b-stt-v2

Add minimal model card (matches v1 style; 5.449% macro WER on ESB)

fix: defensive compat for transformers >=4.55 (PR #135 / @ebezzam repro)

Add higgs_audio_collator.py for run_eval_higgs_audio.py compatibility (parity with v3-8b-stt)

Add transcribe.py for run_eval_higgs_audio.py compatibility (parity with v3-8b-stt)

Initial: chk-38000 + idea-02e83b LoRA merged. Macro WER 5.449% on Open ASR Leaderboard 8 datasets.

initial commit