Hierarchical ASR Model for Speech Recognition Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition technologies face challenges in improving the accuracy of automatic speech recognition (ASR) systems, particularly in translating speech into text effectively, due to limitations in processing audio signals hierarchically and recognizing phonetic features and linguistic information.
Innovation Solution
The implementation of a hierarchical speech recognition model that uses multiple stages of ASR models, including encoders and decoders with projection layers and soft max layers, to process audio signals and output character strings at different levels, such as grapheme, BPE, and word units, enhancing the recognition of speech by integrating acoustic and linguistic information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a single-stage ASR model is used, then the device complexity is low, but the speech recognition accuracy is insufficient
Solution Approach 1:
The ASR model is segmented into multiple stages: a first ASR model that processes audio signals to generate intermediate character strings, and a second ASR model that processes these intermediate results to generate final character strings. This multi-stage segmentation allows each model to focus on specific recognition tasks, improving overall accuracy while managing complexity through modular architecture
Solution Approach 2:
The patent introduces a hierarchical dimension to the ASR process by processing speech recognition at multiple levels: phoneme-level recognition in the first stage and word-level or sentence-level recognition in the second stage. This dimensional expansion from single-level to multi-level processing enables more comprehensive speech analysis and improved recognition accuracy
2Measurement precision
If hierarchical processing with multiple ASR models is implemented, then speech recognition accuracy is improved, but the processing time increases
Solution Approach 1:
The first ASR model performs preliminary processing of audio signals to generate intermediate character strings that capture essential phonetic and linguistic features. This preliminary action prepares the data in advance for the second ASR model, reducing the computational burden during final recognition and optimizing overall processing efficiency
Solution Approach 2:
The patent establishes a continuous processing pipeline where the output of the first ASR model directly feeds into the second ASR model without interruption. This continuous action ensures that useful information is preserved and transformed efficiently through each stage, maintaining high accuracy while minimizing idle time between processing steps
Data Source
AI summary
Provided are an electronic device for recognizing speech of a user, and a method, performed by the electronic device, of recognizing speech. The method includes obtaining an audio signal based on a speech input based on the audio signal being input, obtaining an output value of a first automatic speech recognition (ASR) model that outputs a character string at a first level; obtaining an output value of a second ASR model that outputs a character string at a second level corresponding to the audio signal based on the output value of the first ASR model based on the audio signal being input; and recognizing the speech from the output value of the second ASR model.