Speech Recognition Biasing Using Voice Query History
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech recognition systems struggle to accurately transcribe unique phrases due to the feedback loop issues posed by misrecognized voice queries, leading to inaccurate transcriptions and loss of user trust.
Innovation Solution
A biasing context model is trained using voice query history data to generate a biasing context vector, which augments the lattice of candidate hypotheses, enhancing the likelihood scores of n-grams that are more likely to be spoken by the user, thereby improving transcription accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If speech recognition systems use standard transcription methods without user-specific biasing, then the system maintains simplicity and general applicability, but transcription accuracy for unique phrases deteriorates due to feedback loop issues from misrecognized queries
Solution Approach 1:
The system performs preliminary actions by collecting and storing voice query history data before transcription tasks. A biasing context model is trained in advance on this historical data to learn user-specific n-gram patterns. During actual transcription, this pre-trained model generates biasing context vectors that guide the speech recognizer toward more accurate hypotheses, resolving the contradiction by preparing user-specific information beforehand rather than processing it in real-time
Solution Approach 2:
The patent introduces a biasing context model as an intermediary component between the audio input and the speech recognizer. This model processes voice query history data and generates biasing context vectors that mediate the recognition process, influencing the lattice of candidate hypotheses without replacing the core speech recognition system. This intermediary approach improves accuracy for unique phrases while maintaining the existing system architecture
Solution Approach 3:
The system changes parameters by incorporating biasing context vectors that modify the probability distributions over candidate hypotheses. The biasing context model adjusts the likelihood scores of different n-grams based on user-specific patterns learned from history, dynamically changing the recognition parameters to favor user-specific terminology and phrases, thereby improving transcription accuracy without fundamentally altering the system structure
2Measurement precision
If the system collects and processes extensive voice query history data, then user-specific transcription accuracy improves, but data processing time and computational resources increase
Solution Approach 1:
The biasing context model is trained on voice query history data in advance, before actual transcription tasks. This preliminary training phase consolidates user-specific patterns into a compact model structure. During live transcription, the system only needs to generate biasing context vectors from this pre-trained model, significantly reducing real-time processing time while maintaining high transcription accuracy
Solution Approach 2:
The system uses n-grams of varying lengths (bigrams, trigrams, and longer sequences) to capture user-specific patterns at different levels of granularity. By incorporating multiple n-gram orders and selectively applying biasing based on confidence thresholds, the system achieves high accuracy without processing every possible historical query, balancing thoroughness with efficiency
Data Source
AI summary
A method of using voice query history to improve speech recognition includes receiving audio data corresponding to a current query spoken by a user and processing the audio data to generate a lattice of candidate hypotheses. The method also includes obtaining voice query history data associated with the user that includes n-grams extracted from transcriptions of previous queries spoken by the user, and generating, using a biasing context model configured to receive the voice query history data, a biasing context vector. The biasing context vector indicates a likelihood that each n-gram from the n-grams extracted from the transcriptions of the previous queries spoken by the user will appear in the current query. The method also includes augmenting the lattice of candidate hypotheses based on the biasing context vector and determining a transcription for the current query based on the augmented lattice of candidate hypotheses.