Contextual Biasing for Speech Recognition Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional automatic speech recognition (ASR) systems struggle to accurately recognize domain-specific or personalized words and phrases due to their training on general speech data, leading to low language model scores for out-of-vocabulary (OOV) words, which are not static and lack contextual information.
Innovation Solution
The ASR system dynamically incorporates contextual information during the inference stage by assigning bias scores to context terms, pre-computing these scores during the training stage, and distributing them to suppress false triggering errors without constructing an explicit contextual language model or using metadata, thus improving accuracy and robustness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the language model is trained on general speech data, then the ASR system achieves good performance on general transcription tasks, but it fails to accurately recognize domain-specific or personalized words and phrases
Solution Approach 1:
The patent pre-computes bias scores for all context terms during the training stage and stores them in a lookup table. This preliminary action allows the system to quickly apply contextual biasing during inference without performing complex computations in real-time, thereby improving adaptability to domain-specific words while maintaining general transcription accuracy
Solution Approach 2:
The patent applies different treatment to different words by computing bias scores specifically for context terms that appear in the input context. This local quality approach allows the system to enhance recognition of domain-specific words only where needed (in the context) while leaving general words unchanged, thus improving adaptability without compromising general transcription performance
2Measurement precision
If the ASR system dynamically incorporates contextual information by constructing an explicit contextual language model, then accuracy for OOV words improves, but computational overhead increases significantly
Solution Approach 1:
The patent extracts only the necessary contextual information (bias scores for context terms) from the full contextual language model construction process. Instead of building a complete contextual model, it selectively applies pre-computed bias scores to context terms, thereby improving OOV word accuracy while avoiding the computational burden of full model construction
Solution Approach 2:
The patent performs the computationally intensive task of computing bias scores during the training stage and stores them in advance. During inference, the system only needs to perform simple lookups and additions, which dramatically reduces computational overhead while maintaining high accuracy for OOV words
3Measurement precision
If the ASR system uses metadata or class annotations for context terms, then contextual information can be effectively incorporated, but the system complexity and requirements increase
Solution Approach 1:
The patent enables the language model to automatically incorporate contextual information using only the input context itself, without requiring external metadata or class annotations. The system self-serves by computing bias scores based solely on the context terms provided in the input, thereby achieving effective contextual incorporation while eliminating additional complexity requirements
4Measurement precision
If the ASR system applies bias scores to context terms, then accuracy for personalized words improves, but false triggering errors may occur
Solution Approach 1:
The patent introduces a temperature parameter that controls the strength of bias score application. By adjusting this parameter, the system can balance between improving personalized word accuracy and reducing false triggering errors, allowing flexible adaptation to different operational requirements
Solution Approach 2:
The patent incorporates feedback mechanisms that allow the system to learn from false triggering errors and adjust bias score application accordingly. This feedback loop enables the system to improve personalized word recognition while progressively reducing false positives through iterative optimization
Data Source
AI summary
An automatic speech recognition (ASR) system that determines a textual representation of a word from a word spoken in a natural language is provided. The ASR system uses an acoustic model, a language model, and a decoder. When the ASR system receives a spoken word, the acoustic model generates word candidates for the spoken word. The language model determines an n-gram score for each word candidate. The n-gram score includes a base score and a bias score. The bias score is based on a logarithmic probability of the word candidate, where the logarithmic probability is derived using a class-based language model where the words are clustered into non-overlapping clusters according to word statistics. The decoder decodes a textual representation of the spoken word from the word candidates and the corresponding n-gram score for each word candidate.