You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Explain your intended use. Every request is reviewed manually.

Log in or Sign Up to review the conditions and access this dataset content.

Loic Librosa SNR audit

This manually gated dataset contains all 33 transcript chunks from @loicsumforissuchaninterestingcharacter😅👌🏽#streetballbanter#footballfans_QRBjkUAn1mQ. The audio column contains the original 48 kHz mono 96 kbps MP3 chunks and the text column contains the aligned transcript.

Full-source heuristic result

  • SNR: 15.5324 dB
  • Mean spectral flatness: 0.000013
  • Noise floor: 0.1173024625
  • Peak frame RMS: 0.7013385296
  • Classified noisy: true

Calculation

Every source/chunk is independently decoded with librosa.load(path, sr=None). The score exactly follows the requested heuristic: default-frame RMS, default spectral flatness, 10th-percentile RMS as the noise floor, maximum RMS as the peak, and 10*log10(peak^2/(floor^2+1e-10)). librosa_is_noisy is true when SNR is below 20 dB or mean flatness is above 0.05.

This estimates dynamic range relative to quiet frames; it is not laboratory SNR against a known clean reference.

Downloads last month
4