Enes Mahmut Kulak and Cenk Demiroglu, "Diverse Members, Plain Votes: Fine-Tuning and Combining Five Recognisers for Dysarthric Speech in SAPC2 Track 1"
Abstract
We describe our system for Track 1 (unconstrained ASR) of the second Speech Accessibility Project Challenge. We fine-tuned five recognisers from four model families (Parakeet-TDT 0.6B and 1.1B, Qwen3-ASR-1.7B, Whisper large-v2, Cohere Transcribe) and combined them with equal-weight character-level ROVER and a gated cross-utterance post-processing step. Across our scored submissions the Test1 character error rate fell from 12.95% to 8.22%, 5.3 points below the strongest official baseline; the final submissions are not yet scored. We report what each method was worth on a holdout disjoint from Train in speakers and text: freezing the transducer’s decoder and joint network, model soups, natural-text targets, vote weighting, test-time adaptation, and audio-LLM error correction. Member diversity mattered more than member accuracy. As a corrector, Qwen3-Omni-30B-A3B beat every single model, and gated against the vote it beat the vote too, while training on half of the audio.