ASR paragraph breaks

A 270M-parameter Gemma 3 fine-tune that adds paragraph breaks to flat speech-to-text output and changes nothing else. Built for dictation pipelines: it runs after the ASR pass (here, a Parakeet fine-tune), which returns punctuated, cased text as one block. It is sized to run on a phone.

Break F1 0.78 on 26 held-out, hand-reviewed voice memos (precision 0.84, recall 0.72), against 0.00 for the untrained base model; 2 of 26 outputs fell back to the input. Every output is checked: if anything other than whitespace changed, the input is returned instead. The examples are held-out synthetic memos the model never saw. Runs on CPU, so allow 20-60 seconds.

Model · ONNX · Dataset · Code · Email key-line bolding demo