Design and Implementation of an MLOps Pipeline for Audio-Based Crowd Sentiment Analysis in Stadiums

This thesis presents the design, implementation, and evaluation of GRASS, a near real-time MLOps pipeline for crowd sentiment detection from audio. The system addresses the challenge of inferring collective audience sentiment from ambient audio signals, a task relevant to atmosphere monitoring and crowd management in football stadiums. Developed within a Design Science Research (DSR) framework, the pipeline integrates a C++/FFTW backend for sliding-window segmentation, spectrogram generation, and a ResNet-18 classifier, orchestrated through an MLOps stack of Dagster, MLflow, MinIO, InfluxDB, and Grafana, with human-in-the-loop retraining supported via Label Studio and FastAPI. On the Emotional Crowd Sound Dataset (ECSD), GRASS achieved an F1-score of 96%. In addition, a real-time streaming experiment using a laptop microphone demonstrated average per-segment inference latency of 143.6 ms and an end-to-end response time of 336.9 ms per 2 s audio clip, confirming that the pipeline processes input well within near real-time requirements. The scope of validation remains limited. ECSD is proxy data that does not capture the acoustic and cultural complexity of football crowds, and the streaming experiment was restricted to a single-microphone local setup, not a full stadium deployment. As a result, the accuracy of sentiment detection in authentic football stadium environments remains unresolved. The contribution of this thesis is therefore architectural and methodological: the realization of a reproducible, extensible pipeline and its validation as a proof-of-concept on available data. Future work must focus on testing with authentic stadium recordings, scaling to multi-microphone deployments, and embedding ethical safeguards to ensure responsible deployment.