I am a Computer Science student working on a project to analyze dog barks and classify the underlying emotion. I am considering two different architectural approaches for the audio processing pipeline: Two-stage pipeline: An initial model first classifies the dog's breed (or size category), and its output is passed as an input feature (or used to route the input) to a secondary model that predicts the emotional state. Single end-to-end model: A single model trained directly on the audio features to predict emotion, relying on the network to implicitly learn features related to pitch, vocal tract size, and resonance. From a machine learning and audio signal processing perspective (e.g., using MFCCs or spectrograms): Is there precedent or theoretical justification for splitting acoustic classification into a hierarchical/two-stage pipeline based on physical speaker characteristics (like size or breed)? What are the trade-offs regarding error propagation between stages versus the capacity of a single model to generalize across widely varying acoustic features? Any insights into standard practices for multi-stage audio classification architecture would be appreciated.

Full article content could not be extracted automatically. Read the original below.