Sergio Gastón Burdisso

A discriminative adaptation that reads the final prompt token's hidden state through a classification head, producing a label in one forward pass without modifying the backbone, gives a controlled comparison of generative and discriminative inference in an otherwise identical speechLLM.

Encoder models hold their ground against generative LLMs in ASR evaluation, but the latter enhance interpretability and hypothesis selection.

Reinforcement learning outperforms supervised fine-tuning in adapting ASR systems to synthetic speech, achieving a 40% reduction in word error rates.

LLMs can judge speech recognition quality with near-human accuracy, blowing away traditional metrics like Word Error Rate.

Just 4 hours of speech data closes the modality gap in LLM-based ASR, rivaling full-dataset fine-tuning and unlocking effective domain adaptation.

Open-weight models can now generate realistic, long-form doctor-patient conversations with corresponding SOAP notes, providing a valuable resource for training and evaluating long-context audio reasoning systems.

LLM-based ASR can get a context boost without the compute cost: compress prior audio turns into learned latent tokens and retain transcripts to recover accuracy while shrinking the audio footprint.