Deep Feedforward Neural Network for ASR Hypothesis Ranking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech recognition systems face challenges in accurately selecting the best speech recognition result from multiple engines, as low-level combinations of outputs do not preserve high-level linguistic information, and determining which result to select from multiple engines is complex.
Innovation Solution
A framework that uses a deep feedforward neural network to rank multiple hypotheses generated by ASR engines, incorporating NLU-related features and jointly training with intent detection to improve the ranking process, outputting the top-ranked hypothesis along with NLU results such as intent detection and slot filling results.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple ASR engines are used to generate hypotheses, then speech recognition accuracy is improved, but the complexity of selecting the best result increases
Solution Approach 1:
The patent merges multiple ASR engine outputs into a unified hypothesis ranking framework. Instead of separately processing each engine's results and then attempting to select the best one, the system combines all hypotheses from multiple engines into a single ranked list using a deep feedforward neural network that evaluates all hypotheses simultaneously, reducing selection complexity while maintaining accuracy improvements.
Solution Approach 2:
The patent introduces an intermediary ranking framework that acts as a mediator between multiple ASR engines and the final speech recognition result. This intermediary system processes all hypotheses through a unified neural network model that incorporates NLU knowledge, providing a systematic approach to select the best result without manually managing the complexity of multiple engine outputs.
2Productivity
If low-level combinations of ASR outputs are used, then processing efficiency is improved, but high-level linguistic information is lost
Solution Approach 1:
The patent transitions from low-level feature combination to high-level hypothesis ranking by introducing a new dimension of processing. Instead of combining acoustic features or word-level predictions from multiple engines, the system operates at the hypothesis level, evaluating complete speech recognition results using NLU knowledge and semantic understanding, thereby preserving linguistic information while maintaining efficiency.
Solution Approach 2:
The patent changes the parameter level at which multiple ASR outputs are combined. Rather than combining low-level parameters such as phoneme probabilities or word scores, the system evaluates complete hypotheses using high-level parameters including semantic consistency, intent alignment, and slot filling coherence, thus preserving linguistic information while achieving efficient processing.
3Measurement precision
If NLU knowledge is integrated into hypothesis ranking, then speech recognition accuracy is improved, but computational complexity increases
Solution Approach 1:
The patent creates a universal ranking framework that simultaneously performs multiple functions: evaluating acoustic quality, assessing linguistic consistency, and determining semantic appropriateness. The deep feedforward neural network is designed to handle multiple types of features and evaluation criteria in a single unified model, improving accuracy while managing computational complexity through efficient architecture design.
Solution Approach 2:
The patent applies preliminary filtering and feature extraction before the main ranking computation. By pre-processing hypotheses to extract relevant NLU features and organize them in a standardized format, the system reduces the computational burden of the main ranking operation, allowing NLU knowledge integration to improve accuracy without excessive computational overhead.
Data Source
AI summary
A framework ranks multiple hypotheses generated by one or more ASR engines for each input speech utterance. The framework jointly implements ASR improvement and NLU. It makes use of NLU related knowledge to facilitate the ranking of competing hypotheses, and outputs the top-ranked hypothesis as the improved ASR result together with the NLU results of the speech utterance. The NLU results include intent detection results and the slot filling results.