Contextual Speech Recognition Biasing for Rare Word Pronunciation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional ASR systems struggle to accurately recognize words that are rare or have unusual pronunciations due to limited training data and inefficient use of large training sets, especially in end-to-end models, leading to challenges in recognizing proper names and words that do not follow typical pronunciation rules.
Innovation Solution
Incorporating contextual biasing into neural network-based ASR models by injecting both biasing context and pronunciation information, using grapheme and phoneme data derived from relevant phrases, to improve recognition accuracy for rare and unusual words.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional ASR systems use separate acoustic, pronunciation, and language models, then they can achieve reasonable recognition accuracy, but the system complexity increases and training efficiency decreases
Solution Approach 1:
The patent combines separate acoustic models, pronunciation models, and language models into a unified end-to-end speech recognition model. This integration reduces system complexity by eliminating the need for multiple separate models while maintaining recognition accuracy through joint training of all components within a single neural network architecture.
Solution Approach 2:
The unified end-to-end model performs multiple functions simultaneously - acoustic feature extraction, pronunciation prediction, and language modeling - all within a single neural network. This multi-functional approach replaces the traditional multi-model system, reducing complexity while achieving comprehensive speech recognition capabilities.
2Productivity
If end-to-end models use limited training data, then training efficiency improves, but recognition accuracy for rare words deteriorates
Solution Approach 1:
The patent applies contextual biasing before the speech recognition process by injecting context information and pronunciation data into the model in advance. This preliminary preparation enables the model to better handle rare words during recognition without requiring extensive additional training data, thus maintaining training efficiency while improving accuracy for uncommon vocabulary.
Solution Approach 2:
The patent introduces context information and pronunciation data as intermediary elements that bridge the gap between limited training data and the need for accurate rare word recognition. These intermediaries provide additional guidance during the recognition process, enabling the model to accurately transcribe rare words without requiring large amounts of specialized training data.
3Ease of operation
If contextual biasing uses only written representations of phrases, then processing simplicity is maintained, but recognition accuracy for unusual pronunciations deteriorates
Solution Approach 1:
The patent combines multiple types of data - written representations (graphemes) and pronunciation data (phonemes) - into a composite contextual biasing input. This composite approach allows the model to leverage both visual spelling information and auditory pronunciation information, improving accuracy for unusual pronunciations while maintaining processing efficiency through unified input handling.
Solution Approach 2:
The patent adds a pronunciation dimension to the contextual biasing input by incorporating phoneme data alongside traditional grapheme information. This dimensional expansion enables the model to capture pronunciation patterns that are not evident from spelling alone, significantly improving recognition accuracy for words with unusual or non-standard pronunciations.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A method (500) of biasing speech recognition includes receiving audio data (108) encoding an utterance (104) and obtaining a set of one or more biasing phrases (116) corresponding to a context (202) of the utterance. Each biasing phrase in the set of one or more biasing phrases includes one or more words. The method also includes processing, using a speech recognition model (300), acoustic features (109) derived from the audio data and grapheme and phoneme data (118, 120) derived from the set of one or more biasing phrases to generate an output of the speech recognition model. The method also includes determining a transcription (146) for tire utterance based on the output of the speech recognition model.