Skip to content

refactor: replace custom SNR calculator with librosa

Summary

app/utils/snr_frequency.py (473 lines) computes Signal-to-Noise Ratio by launching ffmpeg subprocesses and parsing stdout with regex across three fallback methods. This approach is fragile, hard to test, and tightly coupled to ffmpeg's output format.

Problem

  • Three separate SNR methods all rely on parsing unstructured ffmpeg text output (e.g., astats filter, silencedetect filter, volumedetect filter)
  • Any ffmpeg version upgrade can silently break parsing without raising an exception
  • The fallback chain masks failures — a wrong result is returned instead of an error
  • Unit testing requires mocking subprocess calls, not the actual audio logic
  • 473 lines of custom signal-processing code that the team must maintain indefinitely

Proposed Solution

Replace with librosa (Apache 2.0), the de-facto Python audio analysis library used across academia and industry.

import librosa
import numpy as np

y, sr = librosa.load(path, sr=None, mono=True)
rms = librosa.feature.rms(y=y)[0]
snr_db = 10 * np.log10(np.max(rms) / (np.min(rms) + 1e-10))

librosa handles MP3, WAV, M4A, OGG natively via soundfile/audioread — no subprocess required.

Library signals:

  • License: Apache 2.0
  • Weekly downloads: 4M+
  • Maintenance: Active (NumFOCUS affiliated)
  • Test suite: Comprehensive CI
  • CVEs: None known

Impact

Dimension Current After
LOC 473 custom lines ~20 lines
Fragility Breaks on ffmpeg output changes Pure Python, version-stable
Testability Requires subprocess mocking Standard unit tests
Accuracy Heuristic bounds (5–60 dB hardcoded) Mathematically correct
Dependencies ffmpeg (system) + regex parsing librosa (pip)

Files Affected

  • app/utils/snr_frequency.py — replace entirely
  • Any callers in app/tasks/data_analysis.py and app/tasks/file_processing.py — update import and call site only