IVR Voice Detection Using Mel-Spectrogram CRNN Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Interactive voice response (IVR) voices in audio data degrade the accuracy of voice-related services and are difficult to detect using traditional silence, noise, and music-based voice activity detection methods, as they often contain actual human voices.
Innovation Solution
A method using a pre-trained neural network model, specifically a convolutional recurrent neural network (CRNN) model, to generate mel-spectrogram frames from audio data, extract features with a convolutional neural network (CNN) and long short-term memory (LSTM) models, and determine the presence of IVR voices by calculating probabilities for each frame, allowing for the identification and elimination of IVR voices from audio data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional silence, noise, and music-based VAD methods are used, then the detection process is simple, but the accuracy of detecting IVR voices is poor because IVR voices contain actual human voices
Solution Approach 1:
The patent transforms audio data into mel-spectrogram images, changing the parameter representation from time-domain audio signals to frequency-domain visual representations. This parameter transformation enables the application of computer vision techniques (CNN) to audio analysis, significantly improving IVR voice detection accuracy while maintaining manageable system complexity through the use of established deep learning architectures.
2Measurement precision
If CRNN model with CNN and LSTM is used, then the detection accuracy reaches 99.59% precision and 96.60% recall, but the computational complexity and processing time increase
Solution Approach 1:
The patent applies preliminary action by pre-processing audio data into mel-spectrogram images before analysis. This transformation prepares the data in an optimized format that leverages the strengths of both CNN (for spatial feature extraction) and LSTM (for temporal sequence modeling), enabling high detection accuracy while reducing the computational burden during the actual detection phase by having the complex transformation done beforehand.
3Measurement precision
If IVR voices are not detected and removed, then the audio data processing is faster, but the speech-to-text recognition accuracy degrades due to noise from automated responses
Solution Approach 1:
The patent applies the extraction principle by identifying and separating IVR voices from mixed audio data containing both human speech and automated responses. The CRNN model detects IVR segments, which are then extracted and removed from the audio stream, leaving only clean human speech for speech-to-text processing. This extraction improves recognition accuracy while the efficient mel-spectrogram approach minimizes the processing overhead.
Data Source
AI summary
Disclosed is a method for detecting a voice from audio data, performed by a computing device according to an exemplary embodiment of the present disclosure. The method includes obtaining audio data; generating image data based on a spectrum of the obtained audio data; analyzing the generated image data by utilizing a pre-trained neural network model; and determining whether an automated response system (ARS) voice is included in the audio data, based on the analysis of the image data.