Text Similarity Approaches

What I worked on

Explored ways to measure text similarity and handle transcription confidence in ASR systems. Looked at multiple-model comparison and top-k predictions.

What I noticed

• WER or statistical comparison can check output consistency
• A decoder can return top-k alternatives with probabilities
• Low-confidence predictions could output “unknown”
• Phoneme errors like “low fever” vs “no fever” matter clinically

”Aha” Moment

n/a

What still feels messy

Still unclear which metric (WER vs probability threshold) provides the most reliable safety signal.

Next step

Prototype a simple top-k decoder and compare its confidence handling to multi-model WER.