search

SakethVNS / kaldi_training

0 stars 0 forks source link

Readme updates #1

Open nageshnayak opened 1 year ago

nageshnayak commented 1 year ago

Add details of code that should be run for training. Which file and command to run?

Go through this website https://kaldi-asr.org/doc/kaldi_for_dummies.html for initial understanding of file structure used in kaldi

From this https://kaldi-asr.org/doc/kaldi_for_dummies.html, what sections does one need to focus on (e.g. Data preparation). Are all steps necessary in that? The details for this are talked about in the next paragraph

In Kaldi Operations are executed in stage wise manner. Before proceeding to training we have to prepare four files(utt2spk, spk2utt, text, wav.scp).

of the Readme. Discuss this first and then provide the link to the above with specific sections that one should focus on.

Where should these folders be created?

Next we need dictionary files that contain lexicon, silence and non silence phones.

How are these obtained?

After creating data directory and dictionary directory we can proceed to create language model using dictionary.

How?

Add a config file with whatever parameters are available that you use for training (Data Folder for training, Within each step (MFCC feature extraction, Monophone, Triphone there are multiple parameters that are configurable). What was used and why? How does this change with different data and language?

For someone looking at this code to begin training from scratch for a new language will be extremely difficult. Need to make the steps as simple as possible.

nageshnayak commented 1 year ago

8kHz data training or New language data like Marathi training will be taken up and Readme + config should include any material necessary for this.