kaldi-asr/kaldi

WSJ setup is a mess

Open

#1,846 opened on Aug 28, 2017

View on GitHub
 (7 comments) (0 reactions) (0 assignees)Shell (15,392 stars) (5,359 forks)batch import
help wantedstale-excludestopped development

Description

[I hit send too soon on this; I'm updating the comment.]

I think the time might have come to create an 's5b' version of the WSJ setup. WSJ is the oldest setup and the local scripts are not up to the standard of clarity that we usually expect. Some specific issues:

  • The dictionary preparation scripts (the larger dictionary) are using some special-purpose scripts that I created a long time ago and probably phonetisaurus would be a better choice.
  • Preparation and cleaning of the language modeling data is done in a way that's mixed up unnecessarily with the dictionary preparation-- better to keep separate things separate.
  • The scripts in local/ need to have much clearer and cleaner interfaces- it needs to be clear what the inputs and output are.
  • I'm not convinced that I like the way the text is normalized. Lower-case is more common than upper-case in ASR setups these days so let's do it lower case, and use <unk> (lower-case) instead of <SPOKEN_NOISE>, which is more standard, and make the phone names lower-case instead of upper-case.

Part of my motivation is that we'll be doing some RNNLM stuff with this setup (since we have example scripts for older setups) and the scripts need to be cleaner. I don't know if anyone has the time and inclination to work on this?

Contributor guide

Tech stack
shell
Domain
machine learningdatatooling
Issue type
refactor
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
5
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
over 1 week
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
stale
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
mostly clear
Prerequisites
Kaldi frameworkWSJ datasetShell scriptingASR basics
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
5
Research direction
Investigate the current WSJ setup scripts in the 'egs/wsj/s5/' directory. Review the specific issues mentioned: dictionary preparation scripts using phonetisaurus, mixed language modeling and dictionary preparation, unclear interfaces in local/ scripts, text normalization choices. Look at comments for additional context. Plan a refactored 's5b' version that separates concerns, uses phonetisaurus, lower case text, and standard <unk> token.