Real-Time Call Drop Prediction Using Audio and Transcript Embeddings
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Interactive voice response systems (IVRs) face challenges in predicting when a caller is likely to drop a voice call before resolving their query, leading to unresolved calls and system inefficiencies.
Innovation Solution
A real-time call monitoring machine learning framework processes interactive call feature data objects, including audio and metadata, to generate a call drop likelihood prediction using audio-based, transcript-based, and event-based embedding data objects, enabling prediction-based actions to prevent call drops.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If real-time call monitoring machine learning framework is implemented to predict call drop likelihood, then call drop prediction accuracy is improved, but system complexity increases
Solution Approach 1:
The system segments the call monitoring process into multiple independent machine learning models: an audio data processing model that analyzes voice characteristics, an audio transcript processing model that processes speech-to-text data, and an event sequence processing model that tracks call events. Each model processes specific feature data objects and generates embedding vectors that are combined to produce the final call drop prediction, allowing complex analysis to be divided into manageable modular components
Solution Approach 2:
The system introduces embedding data objects as intermediary representations between raw input data and final predictions. Audio data objects are transformed into audio embeddings, audio transcripts into transcript embeddings, and event sequences into event embeddings. These embedding vectors serve as standardized intermediate representations that facilitate integration across different processing models and enable the final call drop likelihood prediction without requiring direct complex interactions between all input data types
2Reliability
If multiple machine learning models process audio and metadata data objects, then prediction reliability is improved, but processing time increases
Solution Approach 1:
The system performs preliminary processing of input data by extracting and organizing feature data objects before main prediction processing. Audio data objects are pre-processed to extract relevant acoustic features, metadata objects are structured into standardized formats, and event sequences are pre-sorted and filtered. This preliminary organization of data reduces the computational burden during real-time prediction and enables faster processing while maintaining comprehensive analysis across multiple models
Solution Approach 2:
The machine learning models operate continuously during the voice call, processing incoming audio data and events in real-time rather than analyzing complete calls after they end. The system maintains continuous monitoring streams that update call drop predictions dynamically as new data arrives, allowing for timely interventions while the call is still active and reducing overall processing delays
Data Source
AI summary
Embodiments of the present disclosure provide methods, apparatus, systems, computing devices, and/or computing entities for processing a call drop likelihood prediction for an interactive call data object. In accordance with one embodiment, a method is provided that includes: identifying a group of interactive call feature data objects associated with the interactive call data object that comprises an interactive call audio data object and an interactive call metadata object; processing the call feature data objects using a real-time call monitoring machine learning framework to generate the prediction by: processing the call audio data object using an audio data processing machine learning model to generate an audio-based embedding data object, processing the call audio data object using an audio transcript processing machine learning model to generate a transcript-based embedding data object, and generating the prediction based at least in part on the audio-based and transcript-based embedding data objects and the metadata object.