Speech Recognition for Unregistered Names Using Tagged Phoneme Replacement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
End-to-end (E2E) speech recognition models struggle with recognizing infrequent words like person's names without relearning the entire model, as they lack a dictionary for easy registration.
Innovation Solution
Employ two end-to-end models to tag specific classes like person's names in text and phonemes, followed by phoneme replacement and conversion using a language model to improve recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If an end-to-end model is used for speech recognition, then recognition rate is improved and multi-stage processing is simplified, but the ability to easily register new words like person's names is lost
Solution Approach 1:
The speech recognition system is divided into multiple independent end-to-end models, each specialized for specific vocabulary classes. This segmentation allows each model to be trained and updated independently, enabling easy registration of new words without retraining the entire system.
Solution Approach 2:
A dictionary component is introduced as an intermediary between the speech input and the end-to-end models. The dictionary stores registered vocabulary and routes speech segments to appropriate models, enabling flexible addition of new words through dictionary updates rather than model retraining.
2Device complexity
If a single end-to-end model is used, then device complexity is reduced, but recognition accuracy for specific classes like person's names deteriorates
Solution Approach 1:
The recognition system is segmented into multiple specialized end-to-end models, each optimized for specific vocabulary classes. This segmentation improves recognition accuracy for each class while keeping individual model structures relatively simple.
Solution Approach 2:
The system transitions from a single-model approach to a multi-model architecture, adding the dimension of model specialization. This allows the system to achieve high accuracy for specific classes without requiring overly complex individual models.
3Measurement precision
If relearning is performed for the entire model to recognize new words, then recognition accuracy is maintained, but processing time and computational resources increase
Solution Approach 1:
The system extracts and separates the vocabulary registration function from the main model training process. New words can be added by updating the dictionary and retraining only the relevant end-to-end model, rather than relearning the entire model, significantly reducing processing time.
Solution Approach 2:
Vocabulary is pre-registered in the dictionary before recognition tasks. This preliminary organization allows the system to quickly route speech segments to appropriate models without requiring full model retraining when new words are added.
Data Source
AI summary
A speech recognition device includes: an acquisition part, acquiring a speech signal; a speech feature amount calculation part, calculating a speech feature amount; a first speech recognition part, based on the speech feature amount, performing speech recognition using a learned first E2E model, attaching a first tag to a vocabulary portion of a specific class in text that is a recognition result, and outputting the same; a second speech recognition part, based on the speech feature amount, performing speech recognition using a learned second E2E model, attaching a second tag to a vocabulary portion of a specific class in a phoneme that is a recognition result, and outputting the same; a phoneme replacement part, replacing a vocabulary with the first tag with a phoneme with the second tag; and an output part, converting the phoneme with the second tag into text and outputting the same.