Forget the Recording? Open Speech Models Remember Almost as Well from a Transcript
Year: 2026
Type: preprint
Abstract: A speech assistant that remembers past conversations can keep the old audio or write it down. We ask whether old audio is worth its cost when it competes with text for the same budget of context positions. On VoxMem 8K, MiniCPM-o 4.5 gains nothing measurable from old clips: spending an extra budget of a quarter of the audio's size on them changes accuracy by +0.7 points (95% interval −3.2 to +5.6; 300 question families). For this reader a bare transcript is within 10 points of the full audio history. Gemini 3.8 Flash behaves differently: it scores 18.8 points higher with audio than with a transcript. A text record with speaker, emotion and sound labels brings it to similar accuracy (43.8% against 41.7%), with a wide interval. Code, preregistrations, model outputs and verdicts: https://github.com/SeverinVisionary/voxmem-audio-vs-text-memory AI use: drafted with AI assistance (Claude), reviewed and approved by the author.
Author: Yang, Hanyu
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7223789640