Logo Lanfrica

pdTetteh/ghana-soundscapes-benchmark

Domaine:

environment and energy

Type de record:

dataset
Créateur:
pdT
Hôte:
Benchmark for Ghanaian urban soundscapes with manifests, reproducible baselines, and evaluation workflows for environmental audio classification. # Ghana Soundscapes Benchmark (v1) This repo contains a small environmental audio benchmark built from Ghanaian urban soundscapes (street traffic, generators, trotro transport, markets). It includes: - Raw audio (not in this repo; see “Data” below) - Manifests with standardized metadata - A CNN14-style log-mel baseline model implemented in PyTorch - Reproducible training & evaluation notebook ## Task Multi-class (effectively one-label-per-clip in v1) sound scene classification over 5 Ghanaian acoustic categories: 1. ambient_street_sound – roadside / street ambience 2. generator_hum – power generator noise 3. trotro_engine_sounds – trotro/vehicle engine noise 4. trotro_mate_shouts – mate/assistant vocal shouts in transport 5. market_sounds – market ambience (crowds, sellers, etc.) Each clip is ~N seconds long (you can fill this in) and labeled with one dominant class. ## Data The raw audio is not stored in this repo. Folder structure used locally: ```text Ghana_soundscapes/ data/ ambient_street_sound/ generator_hum/ trotro_engine_sounds/ trotro_mate_shouts/ market_sounds/ manifests/ ghana_soundscapes_raw_wav.csv ghana_soundscapes_metadata_splits_clean.csv ``` ## Repository Structure Suggested layout: ```text Ghana_soundscapes/ notebooks/ ghana_soundscapes_baseline.ipynb scripts/ manifest.py convert_to_wav_and_update_manifest.py manifests/ ghana_soundscapes_raw_wav.csv ghana_soundscapes_metadata_splits_clean.csv environment.yml README.md .gitignore ``` Things that are not committed (but exist locally): ```text Ghana_soundscapes/ data/ ambient_street_sound/ generator_hum/ trotro_engine_sounds/ trotro_mate_shouts/ market_sounds/ runs/ ghana_soundscapes_cnn14/ best_model.pt ... ``` ## Manifests `ghana_soundscapes_raw_wav.csv` Raw manifest built from scanning the `data/` folder. Typical columns: - `filepath` – path to the audio file relative to `data/` - `ext` – file extension (e.g. `wav`) - `format`, `subtype` – audio container/codec info from `soundfi …