Voice Recognition Model Training via Negative Sample Path Constraints

The patent applies preliminary action by constructing negative samples in advance based on positive samples before actual voice recognition. These pre-constructed negative samples are used to train the voice recognition model, enabling it to learn constraints on decoding paths beforehand. This allows the model to avoid unexpected outputs during actual recognition without adding complex real-time constraints during decoding.

The patent implements feedback by using the voice recognition model to generate predictions, then comparing these predictions against the pre-constructed negative samples. The model receives feedback in the form of loss calculations based on whether predictions match expected positive samples or incorrectly match negative samples. This feedback loop continuously refines the model's decoding path constraints during training.

The patent applies local quality by constructing negative samples selectively rather than universally for all possible paths. Instead of creating exhaustive negative samples for every potential decoding path, the method focuses on constructing negative samples based on actual positive samples in the training data. This localized approach creates constraints where they are most needed while avoiding unnecessary complexity in less critical areas.

The patent uses partial action by constructing negative samples only for the specific positive samples present in the training data, rather than attempting to cover all possible voice recognition scenarios. This partial coverage is sufficient to provide effective constraints on decoding paths while keeping the data processing complexity manageable. The approach recognizes that covering all possible paths would be excessive and unnecessary.