Background

The loudness() function in torchaudio.functional implements the ITU-R BS.1770-4 loudness measurement. It applies both absolute and relative gating to compute energy-averaged LKFS values.

Issue

When all audio blocks are below the absolute gating threshold (gamma_abs = -70), the computation of energy_filtered involves division by the count of gated blocks, which is zero in this scenario. This results in a NaN output. This edge case commonly occurs for very quiet or silent audio signals.

Changes

• Introduced torch.where to safely handle cases where the count of gated blocks is zero.
• When no blocks pass the gating threshold, energy_filtered is set to zero.
• Separate variables (abs_gated_blocks and rel_gated_blocks) are used to clearly distinguish absolute and relative gating steps.
• Maintained compatibility with ITU-R BS.1770-4 standard for normal audio signals.

Impact

• Prevents NaN values for extremely low-level signals.
• No change in behavior for signals with blocks above the threshold.
• Safe for both CPU and CUDA devices.