Delhi Street Noise

Street-level noise measurements from 655 locations across Delhi.

Results

We recorded sound meter readings while walking and riding Delhi streets over 20+ days in April-May 2026.

Summary (GoPro continuous + Rider spot measurements):

• 655 locations across Delhi (190 GoPro, 465 Rider)
• 3.4+ hours of continuous observation (12,159 GoPro readings at 1/second)
• Median GoPro noise: 70 dB | Median Rider max: 81 dB
• 4.3% of GoPro readings exceed 85 dB (NIOSH occupational limit)
• 38% of Rider stops recorded peaks ≥85 dB

How noisy is Delhi?

Noise Level	GoPro (% of Time)	Rider (% of Stops)	Reference
≥70 dB	52%	82%	Louder than normal conversation
≥80 dB	11%	52%	Hearing damage risk with prolonged exposure
≥85 dB	4.3%	38%	NIOSH 8-hour occupational limit
≥91 dB	1.4%	25%	NIOSH 2-hour limit

Comparison to Other Cities

How does Delhi compare to urban traffic noise elsewhere? At traffic stops and jams, you're largely hearing engine idling and slow-moving vehicles - conditions that should produce similar noise levels across cities.

Location	Average dB	Peak Range	Source
Delhi (this study)	70 dB (GoPro mean)	49-128 dB	Continuous measurement
NYC streets	73.4 dBA	56-95 dB	Neitzel et al. 2011
NYC high traffic	77.2 dBA	-	Same study
Manhattan	74.6 dBA	-	Same study
US busy urban street	70-80 dB	-	CODOT
EU road traffic (67M exposed)	>55 dB Lden	-	EEA 2024

Key finding: Delhi's average street noise (~70 dB) is comparable to NYC and other busy urban environments. However, Delhi shows more extreme peaks - 38% of rider stops recorded ≥85 dB, and 25% exceeded 91 dB. This suggests Delhi's noise problem is less about average levels and more about frequent high-intensity spikes (horns, engines revving).

Noise by Road Type (Rider Data)

Road Type	N	Mean Max dB	% ≥85 dB
Trunk (major highway)	21	99.5	71%
Primary	30	94.7	70%
Secondary	65	92.3	65%
Tertiary	110	84.7	36%
Residential	194	77.3	18%

Location-level findings (GoPro continuous data)

• 69% of stops (131/190) had at least one reading ≥85 dB
• 12% of stops (23/190) hit peaks ≥100 dB
• Median peak across stops: 89 dB
• Only 1 stop had sustained high noise (>50% of readings ≥85 dB)

Most locations experience brief spikes above safe levels, but sustained harmful exposure is rare.

Rider data findings (465 spot measurements)

• Mean max dB: 84 across all stops
• Mean min dB: 66 (baseline noise level)
• 81 traffic stops and 19 traffic jams recorded
• Road type strongly predicts noise: trunk roads average 99.5 dB max vs 77.3 dB on residential streets

Figures

GoPro analysis (output/analysis/figs/):

• delhi_gopro_fig1_map.html - Interactive map of measurement locations
• delhi_gopro_fig2_histogram.pdf - Distribution of all readings
• delhi_gopro_fig3_stop_distributions.pdf - Per-location summary statistics

Rider analysis (output/rider/analysis/figs/):

• delhi_rider_fig1_map.html - Interactive map of rider stops
• delhi_rider_fig2_histogram.pdf - Distribution of max dB readings
• delhi_rider_fig4_boxplot.pdf - Noise by road type
• delhi_rider_fig5_scatter.pdf - Min vs max dB correlation

Comparison (output/comparison/figs/):

• delhi_compare_fig1_map.html - Map showing both datasets with matches
• delhi_compare_fig2_correlation.pdf - Rider max dB vs GoPro mean dB
• delhi_compare_fig4_road_type.pdf - GoPro noise by road type

Data

Processed Data

File	Description
output/readings/*.json	GoPro OCR'd decibel readings with GPS
output/rider/readings.json	Rider spot measurements with road type metadata
output/analysis/delhi_gopro_data.parquet	GoPro readings as DataFrame
output/rider/analysis/delhi_rider_data.parquet	Rider data with road type
output/comparison/delhi_compare_matched.parquet	GoPro-Rider location matches
output/*/tabs/delhi_*.tex	LaTeX tables

Raw Video (Harvard Dataverse)

Raw GoPro recordings with embedded GPS telemetry:

Archive	Size	Videos
delhi_04_30_2026.tar.gz	1.9 GB	8
delhi_05_01_2026.tar.gz	618 MB	4
delhi_05_03_2026.tar.gz	2.0 GB	18
delhi_05_04_2026.tar.gz	2.0 GB	20
delhi_05_05_2026.tar.gz	2.0 GB	22
delhi_05_06_2026.tar.gz	2.0 GB	24
delhi_05_07_2026.tar.gz	416 MB	3
delhi_05_08_2026.tar.gz	1.9 GB	22
delhi_05_09_2026.tar.gz	1.9 GB	24
delhi_05_13_2026.tar.gz	1.6 GB	27
delhi_05_14_2026.tar.gz	1.9 GB	17
Total	18 GB	192

---

Reproduce the Analysis

Requirements

• Python 3.13+
• ffmpeg, exiftool (brew install ffmpeg exiftool)
• Google API key (for OCR)

Install

git clone https://github.com/soodoku/soundscape
cd soundscape
uv sync
uv pip install -e ".[maps]"

Run Analysis

# Generate figures and tables from existing readings
uv run soundscape analyze --readings output/readings/delhi_full_12728_3-flash-preview_20260523_190001.json

# Open interactive map
open output/analysis/figs/fig1_map_locations.html

Reprocess from Raw Video

If you download the raw videos from Dataverse:

# Extract frames, GPS, metadata
uv run soundscape extract-frames --input data/delhi
uv run soundscape extract-gps --input data/delhi
uv run soundscape extract-exif --input data/delhi
uv run soundscape downsample
uv run soundscape build-manifest

# OCR the sound meter readings
export GOOGLE_API_KEY=your_key
uv run soundscape ocr-readings --model gemini-2.5-flash

# Generate analysis
uv run soundscape analyze --readings output/readings/readings.json

Methods

Data Collection

• Equipment: Handheld sound level meter + GoPro Hero 12 (4K, 50fps, GPS enabled)
• Protocol: Walk to location, hold meter at arm's length, record ~1 minute
• Dates: April 30 - May 14, 2026
• Locations: 190 stops across Delhi (residential, commercial, roadside)

Processing Pipeline

1. Extract frames - Sample 1 frame/second from 50fps video
2. Downsample - Resize to 720p for efficient OCR
3. Extract GPS - Pull coordinates from GoPro telemetry
4. OCR readings - Use Gemini to read decibel value from each frame
5. Filter - Remove readings outside 30-130 dB range (OCR errors)
6. Analyze - Generate tables and figures

OCR Validation

• 95.5% of frames successfully OCR'd
• 5 readings filtered as physically implausible (<30 or >130 dB)
• Manual spot-check via HTML viewer (soundscape viewer)

Pipeline Commands

Command	Description
extract-frames	Extract JPEG frames from video
downsample	Resize frames to 720p
extract-exif	Extract video metadata
extract-gps	Extract GPS from GoPro telemetry
build-manifest	Combine metadata into manifest.json
ocr-readings	OCR sound meter values
viewer	Generate HTML viewer for QC
analyze	Generate GoPro analysis tables and figures
analyze-rider	Generate rider data analysis with road type breakdown
compare-locations	Compare rider and GoPro locations for overlap
create-archives	Create tar.gz for Dataverse

Rider Data Pipeline

For data collected via the rider route tool, conversion scripts are in that repo:

cd ../missing-women-rider-route-tool

# Sound data -> soundscape
python3 export_sound.py \
    --manifest exports/2026-05-24/manifest.json \
    --output ../soundscape/output/rider \
    --geocode

# Pollution data -> streetaqi
python3 export_pollution.py \
    --manifest exports/2026-05-24/manifest.json \
    --output ../streetaqi/data/rider \
    --city delhi \
    --geocode

Analyze Rider Data

# Generate rider-specific analysis (by road type, traffic conditions)
uv run soundscape analyze-rider --readings output/rider/readings.json

# Compare rider and GoPro locations for spatial overlap
uv run soundscape compare-locations \
    --rider output/rider/readings.json \
    --gopro "output/readings/delhi_full_*.json"

License

MIT