Language Model Fusion in ASR

Language model fusion can make a major quality difference, especially in domain-specific speech recognition systems. When acoustically similar but linguistically different candidates exist, an external language model provides a strong guidance signal. This is highly valuable in clinical text, call-center terminology, and proper-name-heavy scenarios. It clearly shows that speech recognition is not only an acoustic decoding problem but also a linguistic decision problem.