Contextual Unidirectional LSTM for Streaming Speech Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning models for speech recognition, particularly unidirectional models, face challenges in processing streaming data due to limitations in temporal dependency, leading to inaccuracies and inability to apply bidirectional models to real-time data streams.
Innovation Solution
The implementation of a contextual unidirectional machine learning model that utilizes LSTM-RNN processing blocks with an embedding vector that incorporates outputs from both temporally preceding and subsequent processing blocks, allowing for a temporal shift in processing flow to enable real-time processing of streaming data without requiring the entire data stream to be received.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If bidirectional RNN models are used to improve accuracy by incorporating future context, then prediction accuracy is improved, but the model cannot be applied to streaming data requiring real-time processing
Solution Approach 1:
The patent applies preliminary action by pre-computing and storing embedding vectors for future frames before they are needed for processing. This allows the unidirectional model to access future context information without actually waiting for or processing future data in real-time, thus maintaining streaming compatibility while achieving bidirectional-like accuracy
Solution Approach 2:
The patent creates copies of future frame data in the form of embedding vectors that are stored in advance. These embedded representations serve as proxies for actual future data, allowing the model to incorporate future context without breaking the unidirectional processing flow required for streaming applications
2Adaptability or versatility
If unidirectional models are used for real-time streaming processing, then applicability to streaming data is improved, but prediction accuracy deteriorates due to lack of future context
Solution Approach 1:
The patent introduces embedding vectors as an intermediary mechanism that bridges the gap between unidirectional processing and bidirectional accuracy. These vectors act as mediators that carry future context information into the unidirectional model without disrupting the forward-only processing flow, thus improving accuracy while maintaining streaming capability
3Measurement precision
If model complexity is increased to improve accuracy through multiple layers and temporal dependencies, then prediction accuracy is improved, but training difficulty increases due to vanishing and exploding gradients
Solution Approach 1:
The patent applies preliminary action by pre-computing embedding vectors offline before the actual streaming processing occurs. This separates the complex computational work from the real-time processing, allowing complex multi-layer models to be trained and prepared in advance without impacting the simplicity and speed of online streaming inference
Data Source
AI summary
Streaming machine learning unidirectional models is facilitated by the use of embedding vectors. Processing blocks in the models apply embedding vectors as input. The embedding vectors utilize context of future data (e.g., data that is temporally offset into the future within a data stream) to improve the accuracy of the outputs generated by the processing blocks. The embedding vectors cause a temporal shift between the outputs of the processing blocks and the inputs to which the outputs correspond. This temporal shift enables the processing blocks to apply the embedding vector inputs from processing blocks that are associated with future data.