CTC Speech Recognition Initialization via HMM State Mapping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Connectionist Temporal Classification Recurrent Neural Networks (CTC-RNNs) in speech recognition often fail to achieve optimal convergence during training, leading to inferior performance compared to traditional Hidden Markov Model (HMM) based systems, due to the lack of a priori frame labeling which results in incomplete and suboptimal alignment of input feature vectors to output phonemes.
Innovation Solution
Initializing a CTC Deep Learning Neural Network for speech recognition using the alignments and frames of an HMM-based speech recognition system, by mapping HMM states to CTC states and training with a cost function to ensure optimal alignment, followed by iterative training and optimization to refine the CTC model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If CTC-RNN is used for speech recognition without a priori frame labeling, then the system achieves greater flexibility and automation, but the alignment of input feature vectors to output phonemes becomes incomplete and suboptimal
Solution Approach 1:
The patent applies preliminary action by using HMM-based speech recognition to generate frame-level phoneme alignments before training the CTC-RNN. These pre-computed alignments serve as training targets, enabling the CTC-RNN to learn optimal mappings without requiring manual a priori labeling, thus resolving the contradiction between automation and alignment precision.
2Manufacturing precision
If traditional HMM-based speech recognition is used, then optimal alignment of phonemes is achieved, but the system lacks the flexibility and modern performance of neural network approaches
Solution Approach 1:
The patent uses HMM-based alignment results as an intermediary to train the CTC-RNN system. The HMM alignments serve as a mediating training signal that transfers the alignment precision advantage to the neural network approach, allowing the CTC-RNN to achieve both high precision and the flexibility of modern deep learning systems.
3Ease of manufacture
If CTC-RNN is trained without pre-computed alignments, then the system is simpler to implement, but training convergence becomes suboptimal and performance deteriorates
Solution Approach 1:
The patent applies preliminary action by computing HMM-based alignments before CTC-RNN training. This pre-computation step provides reliable training targets that ensure optimal convergence, while the overall implementation remains relatively simple by leveraging existing HMM tools and straightforward training pipelines.
Data Source
AI summary
A method for improved initialization of speech recognition system comprises mapping a trained hidden markov model based recognition node network (HMM) to a Connectionist Temporal Classification (CTC) based node label scheme. The central state of each frame in the HMM are mapped to CTC-labeled output nodes and the non-central states of each frame are mapped to CTC-blank nodes to generate a CTC-labeled HMM and each central state represents a phoneme from human speech detected and extracted by a computing device. Next the CTC-labeled HMM is trained using a cost function, wherein the cost function is not part of a CTC cost function. Finally the CTC-labeled HMM is trained using a CTC cost function to produce a CTC node network. The CTC node network may be iteratively trained by repeating the initialization steps.