Study finds language discrimination can narrow the learning gap in multilingual speech modelsMachine translation

In a controlled English–French HuBERT study, researchers improved a bilingual model’s speech and linguistic measures using an auxiliary language classifier or per-language training targets. Phone discrimination error fell from 11.6% to 10.4%. Most measures gained the most when language discrimination was introduced in the first training iteration; later or repeated interventions brought smaller gains and greater separation by language.