Shallow Fusion With Language Models in ASR

A speech recognizer scores possible transcripts from audio. An external language model scores how plausible those word sequences are as text. Shallow fusion combines the scores during search, usually by adding a weighted language-model log score to the recognizer’s score. Research on sequence-to-sequence ASR describes this as a way to incorporate separately trained text knowledge at inference time. The model weights need not be merged or the acoustic model retrained merely to try the combination. Why use it? Some phrases or domain terms are rare in paired audio-training data but more common in available text. The extra language model can guide a beam search toward those sequences when the sound evidence is ambiguous. A beam retains several partial hypotheses instead of committing to the first token. The language-model weight and any length or insertion adjustment affect which completed transcript wins. They should be chosen on development data, leaving a separate test set for evaluation. The method has a failure mode: text plausibility is not evidence that a word was spoken. If the language model is overweighted, it can prefer a fluent domain phrase over an acoustically better alternative. A clinical note that “sounds right” may still misstate a negation or dose. Domain text may also encode biases or privacy-sensitive vocabulary. Measure substitutions, deletions and insertions, examine critical terms and test on both target and general speech. An external language model can improve one slice while hurting another. Shallow fusion is distinct from simply correcting a finished transcript afterward. It influences the search while candidates are still being built. It is also not the only way to combine language information; some ASR models already learn strong language patterns internally. Report the base recognizer, external model, weight, beam settings and evaluation conditions. Treat the final words as an estimate grounded in audio, not as a sentence completion exercise.