Voice Recognition Model Handling Mixed Sampling Rates
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech recognition systems face challenges with high annotation costs and error accumulation due to multiple serial sub-models, and the performance of DNN-based models is limited by the difficulty in obtaining and processing audio data with different sampling rates.
Innovation Solution
A method for speech recognition that involves acquiring first and second linear spectrums from audio data with different sampling rates, performing normalization and Mel-spectrum feature extraction, and training a machine learning model using these features to handle varying sampling rates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If multiple serial sub-models are used for speech recognition, then the system can perform comprehensive speech recognition functions, but annotation cost and time cost increase significantly
Solution Approach 1:
The patent merges multiple serial sub-models (acoustic model, pronunciation dictionary model, language model) into a single unified end-to-end deep neural network model. This integration eliminates the need for separate annotation processes for each sub-model, significantly reducing annotation time cost while maintaining comprehensive speech recognition functionality through the unified model's multiple output layers.
2Reliability
If multiple serial sub-models are used for speech recognition, then the system can perform comprehensive speech recognition functions, but error accumulation occurs across sub-models
Solution Approach 1:
By combining multiple sub-models into a single end-to-end DNN model with shared hidden layers, the patent eliminates error accumulation between separate models. The unified model processes speech input through shared feature extraction layers followed by separate output layers for different recognition tasks, ensuring consistent feature representations and avoiding compounding errors from sequential model processing.
3Adaptability or versatility
If audio data with different sampling rates are used for training, then the model's adaptability improves, but processing complexity increases
Solution Approach 1:
The patent applies parameter changes by dynamically adjusting the FFT size and Mel-filter bank configuration based on the input audio's sampling rate. For different sampling rates (e.g., 16kHz, 32kHz, 44.1kHz), the model automatically configures appropriate spectral analysis parameters, allowing it to process diverse audio data without manual preprocessing while maintaining processing efficiency through parameter adaptation rather than complex architectural changes.
Data Source
Figure 1
Figure 2~3A
Figure 3B~4A
AI summary
A voice recognition method, comprising: acquiring first linear frequency spectrums corresponding to audio to be trained that has different sampling rates; determining the maximum sampling rate and other sampling rates among the different sampling rates; determining the maximum frequency domain sequence number of the first linear frequency spectrums that correspond to the other sampling rates and the maximum sampling rate, respectively, as a first frequency domain sequence number and a second frequency domain sequence number; in the first linear frequency spectrums corresponding to the other sampling rates, configuring an amplitude value corresponding to each frequency domain sequence number that is greater than the first frequency domain sequence number and less than or equal to the second frequency domain sequence number to be zero so as to obtain a second linear frequency spectrum corresponding to the other sampling rates; determining a first voice feature and a second voice feature according to a first Mel spectrum feature of the first linear frequency spectrum corresponding to the maximum sampling rate and a second Mel spectrum feature of the second linear frequency spectrum corresponding to the other sampling rates, respectively; and using the first voice feature and the second voice feature to train a machine learning model. Further disclosed are a voice recognition apparatus and a computer-readable storage medium.