Apple AI Updates: October 3, 2026

1. Apple Found That Teaching Multilingual Speech Models to Tell Languages Apart Closes Most of the Gap to Monolingual Models

Apple Machine Learning Research. Apple researchers trained English/French HuBERT models with two changes meant to sharpen language discrimination: an auxiliary language classifier and per-language k-means targets. Phone discrimination error (phone-ABX) improved from 11.6 percent to 10.4 percent, slightly better than the 10.8 percent monolingual baseline, while lexical scores on sWUGGY rose from 52.1 to 56.7 percent (monolingual 58.5) and prosody scores on ProsAudit rose from 68.9 to 72.9 percent (monolingual 72.6). The largest gains on most measures came when language discrimination was added in the first training iteration, with smaller returns when it was added later. Source

2. Apple Showed That Discrete Diffusion Samplers Cannot Match Dependent Tokens Written in the Same Step

Apple Machine Learning Research. Apple researchers analyzed remasking and uniform-state samplers for discrete diffusion language models, which write several token positions at once from per-position distributions. They proved that a sampling step matches the training distribution only when the positions being written are conditionally independent given the tokens already fixed, and that no product of per-position distributions can reproduce a group of dependent tokens. On ScanAndAdd, a synthetic task with a known closed-form distribution, the generated distribution’s total variation distance was about 29 times the sampling-noise floor. Source