HMM Graph Compatibility Transition to End-to-End Neural Architectures

Kaldi maintains a distinct boundary between legacy HMM-based models and modern end-to-end neural architectures. Each approach requires a specific graph-processing pipeline to ensure that the feature extractor and the decoder are aligned.

When transitioning a deployment from HMM-based decoding to neural architectures, the configuration of the scf file and the associated .fst model files must change to accommodate different feature dimensions and topology requirements. There is uncertainty regarding how to maintain consistent word-error rates when the Viterbi decoder encounters high-noise segments during this transition, as these events often lack explicit error codes in the logs.

• What are the specific configuration changes required in the graph-processing pipeline to move from HMM-based models to end-to-end neural architectures?
• Which verification methods can confirm that the feature-vector alignment is preserved across these different architectural pipelines?