Speech Signal Processing Adjusting Probability Values
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural language models trained only on text data face a mismatch with speech signals, leading to deletion errors and inability to recognize non-verbal sounds, causing premature termination of sentence recognition and inadequate probability calculations.
Innovation Solution
A speech signal processing method that adjusts probability values based on priorities, decreases non-priority values to negative infinity, and uses a combination of neural networks for speech recognition and language models to generate and process non-verbal tokens, ensuring accurate sentence completion and recognition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If a neural language model is trained only with text data, then the model can be trained efficiently using text-only datasets, but the model experiences a mismatch with speech signals causing deletion errors and premature sentence termination
Solution Approach 1:
The patent combines a speech recognition model with a language model to create a hybrid system. The speech recognition model processes speech signals and generates candidate output tokens, while the language model provides probability values for these tokens. This merging allows the system to maintain training efficiency with text-only language models while improving speech recognition accuracy through the integration of speech-specific processing capabilities.
2Adaptability or versatility
If the language model calculates probability values for all candidate output tokens, then comprehensive language understanding is achieved, but the model cannot adequately handle non-verbal sounds such as coughs or hmm sounds
Solution Approach 1:
The patent applies local quality by differentiating the handling of different token types. Non-verbal sound tokens (e.g., cough, hmm) are assigned fixed probability values independent of the language model's predictions, while verbal tokens use the language model's calculated probabilities. This localized differentiation allows the system to maintain comprehensive language understanding for verbal content while specifically addressing the limitations in handling non-verbal sounds.
3Productivity
If the model terminates sentence recognition when a grammatically complete sentence is detected, then processing speed is improved, but deletion errors occur when the actual sentence is not yet finished
Solution Approach 1:
The patent implements feedback mechanisms where the speech recognition model continuously generates candidate output tokens and the language model provides probability assessments. This feedback loop allows the system to monitor sentence completion status dynamically, using the combination of speech recognition confidence and language model probabilities to determine whether to continue or terminate recognition, thereby reducing premature termination errors while maintaining efficient processing.
Data Source
AI summary
A speech signal processing method and apparatus is disclosed. The speech signal processing method includes receiving an input token that is based on a speech signal, calculating first probability values respectively corresponding to candidate output tokens based on the input token, adjusting at least one of the first probability values based on a priority of each of the first probability values, and processing the speech signal based on an adjusted probability value obtained by the adjusting.