Democratic legitimacy rests not only on periodic elections but on ordinary residents having an ongoing voice in the local decisions

that affect them—and the public-comment period of city council meetings is where that voice is most directly exercised. Yet we still lack scalable, direct measures of who actually speaks in these meetings—and for how long—which has limited empirical research on grass-root democracy. We show that recent advances in audio processing and multimodal large language models make such measurement possible at scale. To our knowledge, this is the first study to measure public participation in local government directly from meeting audio. We develop and validate an audio-based pipeline that operates without transcripts or external identity records. The pipeline first applies speaker diarization to determine who speaks when. It then uses a multimodal large language model to classify each detected speaker as an official or a public participant and calculates measures including public floor-time share. We apply this approach to approximately 2,600 hours of audio collected from 1,565 meetings held in five U.S. cities between 2017 and 2023. Against human annotations, diarization achieves 89.4% accuracy, while speaker-role classification achieves 87.7% accuracy and a public-participant F1 score of 0.832 on 1,761 speakers. The resulting measures reveal that public participants constitute 29.9% of detected speakers but receive a median floor-time share of only 10.0%. Public floor-time share also decreased from 15.5% before 2020 to 4.2% in 2020–2021 and recovered only partially to 9.6% in 2022–2023. These

findings extend previous research showing inequalities in who at- tends local meetings: even when public participants attend, their presence does not translate into proportional voice. Beyond this study, the approach can be scaled to additional cities and adapted to examine finer-grained roles, topics, interactions, and institutional features of public deliberation.