Shared Decoding Layer for Speech Recognition Reducing Power
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition technologies using hidden Markov models require pronunciation analysis, which is inefficient compared to end-to-end speech recognition methods that directly convert speech to text using artificial neural networks, and existing deep learning methods face challenges in improving performance without increasing memory bandwidth and power consumption.
Innovation Solution
A decoding method and apparatus that utilize a shared decoding layer in an artificial neural network for speech recognition, performing two decoding tasks to determine output tokens based on probabilities calculated from both speech information and attention networks, while adjusting synapse weights to optimize performance without external language models, thereby reducing system size and power consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If end-to-end speech recognition using artificial neural networks is used, then speech recognition performance is improved, but memory bandwidth and power consumption increase
Solution Approach 1:
The patent segments the speech recognition task into two distinct decoding tasks performed by separate decoding layers: (1) a first decoding layer that processes speech features and recognized tokens to generate first probabilities, and (2) a second decoding layer that processes only recognized tokens to generate second probabilities. This segmentation allows the system to distribute computational workload and reduce the memory bandwidth requirements compared to a monolithic end-to-end approach, while maintaining improved speech recognition performance.
Solution Approach 2:
The patent dynamically adjusts the synthesis of first and second probabilities to determine the output token. The system can adaptively weight and combine the results from the two decoding layers based on the current recognition context, allowing optimal balance between recognition accuracy and computational efficiency at each decoding step, thereby reducing overall power consumption while maintaining high performance.
2Measurement precision
If end-to-end speech recognition using artificial neural networks is used, then speech recognition performance is improved, but system size increases
Solution Approach 1:
The patent divides the speech recognition system into distinct functional segments: an encoding layer that processes speech input, a first decoding layer that handles speech-feature-based decoding, a second decoding layer that handles token-based decoding, and a synthesis mechanism that combines their outputs. This modular segmentation reduces system size by allowing independent optimization and efficient resource allocation for each component, compared to a single large end-to-end model.
Solution Approach 2:
The patent implements a universal decoding architecture where the same neural network infrastructure supports multiple decoding tasks. The first and second decoding layers share common computational resources and can be configured to perform different functions depending on the input type (speech features vs. recognized tokens), reducing overall system size while maintaining the ability to perform both speech-based and text-based decoding operations.
3Reliability
If pronunciation analysis is performed using hidden Markov models, then speech recognition can be achieved, but recognition efficiency decreases
Solution Approach 1:
The patent extracts and eliminates the pronunciation analysis step from the speech recognition pipeline. Instead of using hidden Markov models to analyze pronunciations before recognition, the system directly processes speech features through the encoding layer and feeds them into the first decoding layer, which performs recognition based on speech features and recognized tokens. This extraction of the intermediate pronunciation analysis step significantly improves recognition efficiency while maintaining reliable speech recognition capability.
Solution Approach 2:
The patent performs preliminary processing of speech input by the encoding layer to extract relevant features before they reach the decoding stages. This preliminary action prepares the data in an optimized format that can be directly used by the first and second decoding layers, eliminating the need for subsequent pronunciation analysis and improving overall recognition efficiency while preserving accurate speech recognition.
Data Source
AI summary
A decoding method and apparatus in an artificial neural network for speech recognition. The decoding method in the artificial neural network for speech recognition includes performing a first decoding task of decoding a feature including speech information and at least one token recognized up to current time, using a shared decoding layer included in the artificial neural network, performing a second decoding task of decoding the at least one token, using the shared decoding layer, and determining an output token to be recognized subsequent to the at least one token based on a result of the first decoding task and a result of the second decoding task.