Skip to content

Latest commit

 

History

History
26 lines (20 loc) · 1.75 KB

README.md

File metadata and controls

26 lines (20 loc) · 1.75 KB
Raw

Speech Recognition Tools

This repository contains miscellaneous tools for Automatic Speech Recognition(ASR) based on Kaldi or ESPNet

To use ESPnet recipes

  • Install ESPnet (https://github.com/espnet/espnet) under speech_recognition_tools/tools/.
  • If you already have ESPnet installed you can provide a symbolic link to it

To use Kaldi recipes

  • Install my Kaldi fork (https://github.com/sadhusamik/kaldi) under speech_recognition_tools/tools/
  • You can use any newer versions of Kaldi but might have to change some scripts
  • If you already have Kaldi installed you can provide a symbolic link to it

FDLP-spectrogram (with ESPnet)

  • Paper: (https://arxiv.org/abs/2103.14129)
  • The WSJ and REVERB recipes for FDLP-spectrogram can be found under e2e/wsj and e2e/reverb.
  • The script run_fdlp_e1.sh runs the standard ESPnet recipes for each dataset with FDLP-spectrogram features instead of the usual mel-spectrogram (log filter-bank energy)
  • The script run_melspec.sh runs the ESPnet baseline mel-spectrogram features. NOTE: This is different from the Kaldi recipe of mel-spectrogram
  • The script run_fdlp_e1.sh provides the template to use FDLP-spectrogram for other datasets
Data set mel-spectrogram FDLP-spectrogram
WSJ (test_eval92) 5.1 4.8
REVERB (et_real_1ch / et_real_1ch_wpe / et_real_8ch_beamformit) 23.2 / 20.7 / 9.2 19.4 / 18.0 / 7.2
CHIME4 (et05_real_isolated_1ch_track / et05_real_beamformit_2mics / et05_real_beamformit_5mics) 23.7 / 20.4 / 16.8 23.4 / 19.5 / 15.8