Minilm L12 ASR Paragraph Breaks
No description provided
About This Project
MiniLM-L12-ASR-Paragraph-Breaks is a fine-tuned BERT-style token classifier that adds paragraph breaks to punctuated, cased speech recognition output, which often arrives as a single block of text. The model labels each word as either a paragraph start or not, and the output is the original words joined by spaces or blank lines. Because it is a classifier rather than a generator, it cannot change the wording of the transcript. The model is built on microsoft/MiniLM-L12-H384-uncased and was trained on the danielrosehill/ASR-Paragraph-Breaks dataset of synthetic voice memos, with added ASR-style noise such as lowercase sentence starts, dropped commas, and inserted fillers.
The project is distributed in PyTorch safetensors format and as ONNX models, including an int8 dynamically quantized version of about 32 MB intended for application deployment. Inference uses ONNX Runtime with WordPiece tokenization, sliding 510-token windows, and a tuned paragraph threshold. A reference script, paragraphs.py, is provided and can be run with uv. On a held-out set of 26 synthetic memos, the model reported an F1 of 0.85, compared with 0.78 for a generative Gemma 3 270M fine-tune and 0.65 for the SaT segmentation baseline in paragraph mode. Reported int8 latency on a laptop CPU averaged 0.65 seconds for a 340-word memo.
The model is suited to developers building dictation, voice note, or transcription tools who need lightweight, on-device paragraph segmentation after speech recognition. It is English only and was evaluated on a small, single-style test set, so its accuracy on other writing styles is not established. It reads the full text at once and is intended to run after transcription is complete rather than during live dictation.