tf_mask

Time-frequency Mask

Computation of the time-frequency mask (PSM, IRM, IBM, IAM, ...) as the neural network training labels.

Cmd options

See ./scripts/sptk/compute_mask.py -h

Usage

1. IBM & IRM computation

   # prepare scp
   echo "egs asset/clean.wav" > clean.scp
   echo "egs asset/noisy.wav" > noisy.scp
   # computation
   ../../scripts/sptk/compute_mask.py \
       --mask irm clean.scp noisy.scp irm.ark
   # visualize and check
   ../../scripts/sptk/visualize_tf_matrix.py \
       --input ark \
       --cmap jet \
       --cache-dir irm \
       irm.ark

2. PSM & IAM (FFT-mask or SMM) computation

   # add cutoff as they are unbounded
   ../../scripts/sptk/compute_mask.py \
       --mask psm \
       --cutoff 2 \
       clean.scp noisy.scp psm.ark
   # visualize and check
   ../../scripts/sptk/visualize_tf_matrix.py \
       --input ark \
       --cmap jet \
       --cache-dir psm \
       psm.ark

3. Restore audio using TF-masks

   # psm as example
   ../../scripts/sptk/compute_mask.py \
       --mask psm \
       --cutoff 2 \
       --scp mask.scp \
       clean.scp noisy.scp mask.ark
   # do TF masking (using noisy phase)
   ../../scripts/sptk/wav_separate.py \
       --mask-format kaldi \
       noisy.scp mask.scp enh

   The enhancement output is under directory enh. See ../../scripts/sptk/wav_separate.py -h for more command options.