Voicemail Detection via Audio-to-Image ML Conversion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current answering machine/voicemail system detection techniques are inaccurate, often mistaking voice mailbox messages for live persons or incorrectly identifying delays as voicemail systems, leading to inefficiencies in call handling.
Innovation Solution
A method and system using a machine learning model that converts audio files into images, specifically Mel-spectrograms, to predict whether an answering machine/voicemail system is present, with weight adjustments in the model to improve prediction accuracy, and sends indications via text-to-speech messages when an answering machine/voicemail system is detected.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If current answering machine/voicemail system detection techniques are used, then call handling can be automated, but detection accuracy deteriorates due to mistaking voice mailbox messages for live persons or incorrectly identifying delays as voicemail systems
Solution Approach 1:
The patent replaces traditional mechanical rule-based detection systems with a neural network-based machine learning system. The neural network analyzes audio features and patterns to distinguish voicemail systems from live persons, achieving both automation and high accuracy without the false positives that plague rule-based systems.
Solution Approach 2:
The patent transforms audio data into different parameter representations (spectrograms, mel-frequency cepstral coefficients) that reveal distinctive patterns of voicemail systems versus human voices. By changing the parameter space from raw audio waveforms to frequency-time representations, the system can accurately detect voicemail characteristics while maintaining automation.
2Device complexity
If traditional detection methods are used, then system complexity is reduced, but detection reliability deteriorates due to incorrect identifications
Solution Approach 1:
The patent replaces simple but unreliable mechanical detection rules with a neural network system that, while more complex in structure, provides superior reliability through learned patterns. The neural network's ability to generalize from training data makes it more reliable than hand-crafted rules, justifying the increased complexity.
Solution Approach 2:
The patent performs preliminary training of the neural network on labeled voicemail and human voice data before deployment. This preliminary action allows the system to learn reliable detection patterns in advance, ensuring high reliability during actual operation without requiring complex real-time decision logic.
3Measurement precision
If machine learning models with weight adjustments are used, then detection precision is improved, but computational requirements and processing time increase
Solution Approach 1:
The patent performs weight adjustments and model training in advance during a preliminary phase. Once trained, the model can make rapid predictions during actual call handling without requiring real-time weight adjustments. This separates the computationally intensive training phase from the fast inference phase, resolving the time precision trade-off.
Solution Approach 2:
The patent segments the audio stream into manageable frames and processes them through the neural network in stages. By dividing the audio processing into smaller temporal segments and using efficient feature extraction (mel-spectrograms), the system achieves high precision while keeping processing time acceptable for real-time or near-real-time operation.
Data Source
AI summary
Methods and systems for detecting an answering machine/voicemail system using a machine learning model are provided herein. In some embodiments, a method for detecting an answering machine/voicemail system using a machine learning model comprises receiving an audio stream from a telecommunication session; parsing the audio stream into a plurality of audio files; converting each of the plurality of audio files into an image; inputting each of the converted images into the machine learning model; receiving a prediction from the machine learning model; sending an indication that an answering machine/voicemail system is detected when the received prediction is a beep.