sprocket-el

Fork of sprocket speicialized in aligning normal and electrolarynx speech.

How to run

1. Prepare wav files. Use test_prepare_*_files.py to rename, resample to 16 kHz, and copy files from source to the working folder, i.e. under example/data/wav/.
2. Also use test_cut_wavs.py and test_stretch_audio.py to cut initial transients and pre-stretch WAV fiels if needed.
3. Run initialize.py steps 1, 2, and 3.
4. Modify the *.yml and *.list files if needed.
   • Add fake in the YML file if needed.
   • Use the f0 and npow histograms under conf/figure to set reasonable thresholds.

• See example/conf/speaker/nasal_tsai_mhint_20211128_cut[0.10].yml for an example.

5. Run run_sprocket.py steps 1, 2, 3, and 6. The results will be stored in example/data/pair/*/aligned/.

Files

Modified files

• example/src/yml.py,
• example/src/extract_features.py,
• sprocket/speech/feature_extrator.py,
• sprocket/speech/analyzer.py: Add the ability to provide fake f0 (e.g. 100 Hz electrolarynx excitations) and median-filter the extracted f0 by specifying the kernel size.
• sprocket/speech/extfrm.py: Log the number of non-silent frames extracted.
• example/src/estimate_twf_and_jnt.py: Median-filter the power and also save the joint feature vectors in the HDF5 file format. Useful for outputting the aligned WAV files.
• example/run_sprocket.py: Add an extra step 6, which outputs the aligned WAV files. Also disable step 4 & 5.

Added files

• example/file_utils.py: Add utilites to rename and resample files. Used by test_prepare_{nasal, normal}_files.py
• example/src/output_aligned.py: Take the results of sprocket step 3 (aligning) and resynthesize aligned WAVs. Used in sprocket step 6.

Scripts

• example/test_prepare_{nasal, normal}_files.py: Prepare nasal and normal files.
• example/test_stretch_audio.py: Pre-stretch faster speech files (usually normal speech) to roughly match the slower ones. This depends on the pysox package.
• example/test_cut_wavs.py: Cut the initial part of WAV files; useful to remove the initial transients.