Apple Machine Learning Research reported an iterative pseudo-labeling training approach applied to code-switched automatic speech recognition for the first time.[1] The approach comprises three phases of pseudo-label generation from a large unlabeled corpus, two-stage bilingual model training, and iterative improvements, the research said.[1]
Apple Machine Learning Research reported Mix Error Rate reductions of 6.35 percent on the SEAME devman subset and 8.29 percent on the devsge subset.[1] Apple Machine Learning Research also described a lightweight method that detects code-switching directly from speech using frame-level language information in pretrained speech encoders.[2] The method identifies language-selective hidden-state dimensions using monolingual recordings and aggregates their frame-level activations into a code-switching prediction, the research said.[2] Apple Machine Learning Research reported that the Whisper-small encoder achieves a mean F1 of 0.798 across 14 language pairs in CS-FLEURS, compared with 0.776 for a CNN and 0.751 for a HuBERT linear baseline trained for monolingual language identification.[2]