CHANGELOG.md

Changelog

Unreleased

Added

• Audio preprocessing configuration in backend/src/config.py.
• New backend/src/audio_processing.py module for already-separated stems.
• Stem loading and validation for the fixed stem order: drums, bass, vocals, other.
• Stem alignment by trimming all stems to the shortest available length.
• Temporal audio windowing for model context windows.
• Mel spectrogram generation with nnAudio.
• Spectrogram normalization for neural network inputs.
• Tensor generation for model consumption in [Num_Windows, 4, N_MELS, Time_Steps] format.
• CNN2DMultichannel model architecture.
• 2D CNN architecture for four-stem Mel spectrogram input.
• Output layer for four vibration zones.
• Shape validation in the model forward pass.

Notes

• Stem separation is not performed by audio_processing.py.
• Stem separation belongs to another pipeline step before preprocessing.
• The preprocessing module assumes four stems are already available.
• The current stem order is fixed as drums, bass, vocals, other.
• Length alignment trims to the shortest stem so every generated window contains audio from all stems.
• CNN2DMultichannel expects input tensors shaped [Batch_Size, 4, N_MELS, Time_Steps].
• The four model input channels correspond to [drums, bass, vocals, other].
• The four model output values correspond to [Vib_Seat, Vib_Lumbar, Vib_Neck, Vib_Arms].
• The final sigmoid constrains model outputs to [0.0, 1.0].

Future Work

• Demucs integration.
• Streaming audio support.
• Real-time inference support.
• Data augmentation.
• GPU optimization.