Speech Recognition Named-Entity Correction With Acoustic Embeddings
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Speech recognition devices struggle to accurately recognize named entities due to the vast number of possible entities and variations in pronunciation, leading to incorrect selections and reduced accuracy.
Innovation Solution
A speech recognition method and device that utilizes an acoustic embedding model to generate vectors for speech signals, comparing them to stored vectors in a database to correct identified named entities, accounting for variations in pronunciation and adding new entities based on user input.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If all possible named entities are stored in advance in the speech recognition device, then the accuracy of speech recognition is improved, but the device complexity and storage requirements increase significantly
Solution Approach 1:
The patent introduces an acoustic embedding model as an intermediary between the speech signal and the named entity database. This model converts speech signals into acoustic embedding vectors, which are then compared with stored vectors to identify named entities. This intermediary approach allows the system to handle unlimited named entities without storing all possible entities explicitly, thus improving speech recognition accuracy while avoiding the complexity of maintaining a comprehensive database.
Solution Approach 2:
The patent stores acoustic embedding vectors (copies or representations) of named entities in the database rather than storing all possible named entity variations. By storing compact vector representations instead of complete entity data, the system achieves efficient storage and comparison while maintaining high recognition accuracy. This copying approach reduces device complexity while preserving the ability to recognize diverse named entities.
2Device complexity
If the speech recognition device stores a limited number of named entities, then the device complexity is reduced, but the accuracy of speech recognition deteriorates when encountering unseen entities
Solution Approach 1:
The patent transforms the named entity recognition problem from a discrete matching task to a continuous vector space comparison task. By converting both speech signals and named entities into acoustic embedding vectors, the system can measure similarity through distance metrics in the vector space. This parameter transformation allows the device to handle unseen entities by finding the closest matching vector, thereby maintaining high accuracy with limited stored data.
Solution Approach 2:
The acoustic embedding model serves multiple functions: it converts speech signals to vectors, enables similarity comparison, handles unseen entities through approximate matching, and adapts to different pronunciations. This universal approach allows the system to maintain low device complexity while achieving high speech recognition accuracy across diverse scenarios including unseen and differently pronounced named entities.
3Speed
If the speech recognition device uses traditional phoneme-based matching, then the processing speed is fast, but the accuracy deteriorates when users pronounce named entities differently
Solution Approach 1:
The patent changes the matching parameter from discrete phoneme sequences to continuous acoustic embedding vectors. This transformation allows the system to capture subtle pronunciation variations while maintaining efficient vector comparison. The acoustic embedding model processes speech signals to generate vectors that inherently account for pronunciation differences, enabling both fast processing through vector operations and high accuracy in handling diverse pronunciations.
Data Source
AI summary
Provided are a method and device for speech recognition. The speech recognition method includes: receiving a speech signal generated by an utterance of a user; identifying a named entity from the received speech signal; determining a speech signal portion, which corresponds to the identified named entity, from the received speech signal; generating a first acoustic embedding vector corresponding to the speech signal portion, based on an acoustic embedding model; determining a second acoustic embedding vector that is one of a plurality of acoustic embedding vectors corresponding to a plurality of named entities included in an acoustic embedding database (DB), based on distances between the plurality of acoustic embedding vectors and the first acoustic embedding vector; determining a corrected named entity corresponding to the second acoustic embedding vector; and providing a result of speech recognition with respect to the speech signal, based on the corrected named entity.