Wake-Word Training Using Phoneme Matching to Cut Data Recording
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing wake-up word recognition systems face challenges in achieving high accuracy with excessive time consumption and laborious data recording processes, often requiring large amounts of human involvement and prone to errors.
Innovation Solution
A wake-up word recognition training system utilizing a sentence database, phoneme disassembly, analysis, and a deep neural network model to automate the process, reducing human intervention and improving accuracy through phoneme sequence matching and dynamic programming.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional speech recognition training uses marked data and key information, then the training process can be completed, but it consumes excessive time and has low accuracy
Solution Approach 1:
The patent uses existing sentence databases and speech data as templates to generate training data through phoneme substitution rather than collecting new marked data. The system copies phoneme sequences from existing sentences and replaces them to create wake-up word training samples, eliminating time-consuming manual data annotation while maintaining recognition accuracy
Solution Approach 2:
The patent pre-processes sentence databases by segmenting sentences into phoneme sequences and storing them in advance. This preliminary action creates a ready-to-use phoneme library that can be quickly assembled into training data without real-time processing delays, thus reducing training preparation time while ensuring accuracy
2Measurement precision
If large numbers of people record speech data to fine-tune deep learning models, then the hit rate of wake-up word recognition improves, but the process becomes time-consuming and laborious
Solution Approach 1:
The system automatically generates training data by self-assembling phoneme sequences from existing sentence databases without requiring human recorders. The automated phoneme substitution and sentence assembly processes replace manual data collection, maintaining high hit rates while dramatically improving training productivity
Solution Approach 2:
The patent creates a universal phoneme-based training data generation system that can produce training data for any wake-up word by simply substituting phoneme sequences. This multi-functional approach eliminates the need for separate recording sessions for each wake-up word, thereby improving both hit rate consistency and training efficiency across different wake-up words
3Reliability
If speech data of each wake-up word is recorded independently, then the recognition model can be trained, but the process requires repeating operations which is time-consuming and laborious
Solution Approach 1:
The patent segments wake-up words into phoneme sequences and trains the recognition model on phoneme-level features rather than complete wake-up word recordings. This segmentation allows the system to handle different wake-up words through phoneme substitution without repeating the entire recording process, maintaining training completeness while reducing time consumption
Solution Approach 2:
The system performs preliminary phoneme segmentation and sequence assembly for all potential wake-up words using a unified sentence database. This preliminary action creates a reusable phoneme library that serves all wake-up word training needs, eliminating repeated operations while ensuring each model receives complete and appropriate training data
4Quantity of substance
If manual recording processes are used for wake-up word data, then training data can be collected, but the occurrence of human errors increases
Solution Approach 1:
The patent replaces the mechanical human recording process with an automated computer-based phoneme substitution system. The system automatically generates training data by substituting phoneme sequences in existing sentences, eliminating human errors in data collection while maintaining sufficient training data quantity and improving data accuracy
Data Source
AI summary
A wake-up word recognition training system includes: a sentence database, storing a plurality of sentences and a phoneme sequence and a speech signal corresponding to each of the sentences; a phoneme disassembly module, disassembling a wake-up word inputted from the outside to obtain a wake-up word phoneme sequence; a phoneme analysis module, matching the wake-up word phoneme sequence to the sentences and/or phoneme sequences thereof, to obtain wake-up word part-of-speech sentences and non-wake-up word part-of-speech sentences; a sentence classification module, dividing the sentences in the sentence database into the wake-up word part-of-speech sentences and the non-wake-up word part-of-speech sentences according to a comparison result of the phoneme comparison module; and a wake-up word recognition module, obtaining speech signal fragments of the wake-up word and a non-wake-up word according to a phoneme combination of the wake-up word part-of-speech sentences and the non-wake-up word part-of-speech sentences.