microsoft / DNS-Challenge

This repo contains the scripts, models, and required files for the Deep Noise Suppression (DNS) Challenge.
Creative Commons Attribution 4.0 International
1.13k stars 415 forks source link

Unexpected clean and noisy data #118

Open st3inum opened 2 years ago

st3inum commented 2 years ago

My config:

[noisy_speech]

sampling_rate: 48000
audioformat: *.wav
audio_length: 10 
silence_length: 0.2
total_hours: 500
snr_lower: 0
snr_upper: 40
randomize_snr: True
target_level_lower: -35
target_level_upper: -15
total_snrlevels: 5 
clean_activity_threshold: 0.6
noise_activity_threshold: 0.0
fileindex_start: None
fileindex_end: None
is_test_set: False
noise_dir: datasets/noise
speech_dir: datasets/clean
noisy_destination: noisy
clean_destination: clean
noise_destination: noise
log_dir: logs

# Unit tests config
snr_test: True
norm_test: True
sampling_rate_test = True
clipping_test = True

unit_tests_log_dir: ./unittests_logs

For this config, I've noticed some unexpected behavior in one (clean, noisy) pair. noisy file: book_04362_chp_0020_reader_11254_76_colv997yUuw_snr10_fileid_62459.wav clean file(generated): clean_fileid_62459.wav main clean file: book_04362_chp_0020_reader_11254_76.wav

It seems that the generated clean files also contain noisy data. I think this is the reason for upsampling from 16KHz to 48KHz. But don't know the exact reason. What is the main reason for it, and how can I prevent it?