yunzhongfei / kaldi-senan

kaldi-SENAN

kaldi-SENAN is the implementation of speech-enhanced and noise-aware network (SENAN, see the following paper) built on the open-sourced Kaldi toolkit. Example scripts for Aurora-4 task are provided and located at egs/aurora4/proposed.

Hung-Shin Lee, Pin-Yuan Chen, Yu Tsao, and Hsin-Min Wang, "Speech-enhanced and noise-aware networks for robust speech recognition," submitted to Interspeech 2022.

---

Prerequisites

Follow kaldi installation steps and install this project.

Aurora-4 example

1. In stage 8 of run.sh, change command to

# TDNN-F as AM + proposed model

local/chain/tuning/run_tdnn-1a_mtae_mfcc-mfcc-cont_noise-stats.sh

# TDNN-F as AM + SpecAugment + proposed model

local/chain/tuning/run_tdnn-1a_mtae_mfcc-mfcc-cont_noise-stats_specaugment.sh 

# CNN-TDNN-F as AM + proposed model

local/chain/tuning/run_cnn-tdnn-1c_mtae_mfcc-mfcc-cont_noise-stats.sh

# CNN-TDNN-F as AM + SpecAugment + proposed model

local/chain/tuning/run_cnn-tdnn-1c_mtae_mfcc-mfcc-cont_noise-stats_specaugment.sh 

2. The weight for the two output layers can be changed by modifying frame_weight_dae and frame_weight_dspae in run_{tdnn-1a,cnn-tdnn-1c}\_mtae\_*.sh