Feature Request: Automatic Noise Reduction Based on SNR Analysis

[erklu]

Description

Once FFmpeg is implemented as the first step in the pipeline, it can also be used to analyze the signal-to-noise ratio (SNR) of incoming audio files.
If the SNR is below a defined threshold (e.g., 12 dB), an automatic noise reduction step should be triggered.

The goal is to improve audio quality by detecting and reducing noise before transcribing with Whisper.
This will help ensure better transcription results, especially for recordings with background noise or low-quality input.

⸻

Expected Behavior
• Automatic audio quality analysis is performed prior to transcription.
• If the SNR is below a certain threshold, noise reduction is applied.
• A cleaned version of the audio is created and used as input for Whisper instead of the original file.

⸻

Technical Guidelines

1. Measure the SNR (Signal-to-Noise Ratio) to determine if noise reduction is needed.
2. Apply a noise reduction method when SNR is too low, using one of the following:
   • DeepFilterNet
   • RNNoise
   • FFmpeg noise filters
3. For evaluation purposes, consider running post-reduction SNR analysis to verify the improvement.

⸻

Important Considerations
• Test different noise reduction models to find the best balance between audio quality and computational efficiency.
• Decide whether users should have the option to enable/disable noise reduction, or whether it should be fully automatic when triggered by low SNR.
• Current thinking leans toward a fully automatic fallback in “difficult” cases.
• Optimize the pipeline to only run noise reduction when truly necessary, in order to reduce unnecessary processing time and cost.