Skip to main content

Nadira Povey

Search This Blog

Posts

ASR model and HCLG.fst LibriSpeech

on May 10, 2022

Get link
Facebook
Twitter
Pinterest
Email
Other Apps

Tested train_lm.sh with --include-heldout

on May 10, 2022

+

Get link
Facebook
Twitter
Pinterest
Email
Other Apps

What is HCLG.fst?

on May 10, 2022

Get link
Facebook
Twitter
Pinterest
Email
Other Apps

LibriSpeech training

on May 09, 2022

+

Get link
Facebook
Twitter
Pinterest
Email
Other Apps

Install Kaldi: Ubuntu

on May 07, 2022

+

Get link
Facebook
Twitter
Pinterest
Email
Other Apps

Install Kaldi: Red Hat

on May 06, 2022

+

Get link
Facebook
Twitter
Pinterest
Email
Other Apps

Popular Posts

What is HCLG.fst?

Trying to understand what is needed to make HCLG.fst file.

Elementary Quest Program 2011. Application process for the first grade

Step 1: Submitted application. November 2010

Dynamic Data Parallel

The normal way in Pytorch to use multiple GPUs is to use DDP (Dynamic Data Parallel). In Pytorch documentation they have a very good tutorials. There is a certain boiler plate code we should use. We have to invoke in a special way. It will produce several versions of the process and each version will use a different GPU. It will not change the code, it will be only wrapper around the code. Just by having a code base we normally can't say if it is using multiple GPUs. But we can search for words "rank", "world_size" and "ddp". Please look at this project to get an idea about how one might use DDP https://github.com/k2-fsa/icefall/search?q=ddp

End-to-End STT

End-to-end generally refers to things like CTC, transformers, RNN-T. Generally, anything where you can code the entire thing with PyTorch/TensorFlow without C++ code. but again, End to end is not really a meaningful or important category. Generally it's considered trendy to be end to end.

ASR model and HCLG.fst LibriSpeech

Many final models final.mdl in various folders, which one is the final one we need?

"$train_cmd"

"$train_cmd"

Do the lexicon words in dict folder become unigrams?

Typically the unigram probs are just all the words in the dictionary, but it can depend whether the word-list or dictionary was passed into the LM-building script. If nothing was passed in, unigrams might just be whatever was in the training data but it's normal to pass in the word-list. It's easy to check the number of unigrams, it's at the top of the arpa file.

Reconstruct the Original Training Data from N-Grams.

Reconstruct the Original Training Data from N-Grams.

ImportError: No module named posixpath

ImportError: No module named posixpath If you got the python error message: ImportError: No module named posixpath Please run the command below: virtualenv --version If you have the version: virtualenv 20.4.3 that might be your problem. To fix the problem I deleted my old environment and ran commands below pip install virtualenv==16.7.7 virtualenv -p python2 venv source venv/bin/activate Different version of the virtualenv seems can fix the problem. Install all packages that you need. Hope this helped!

Building a Statistical Language Model for ASR with Kaldi Script

Building a Statistical Language Model. the number on the RHS is the "backoff penalty" that says if you have just seen "job good->", how much of the probability mass belongs to the lower-level history, i.e. the state "good->". If there is no such history, e.g. there no n-grams "job arises -> X", there is no need for this number. Example for a 3gram with perplexity numbers on the left. Note 3gram has bigrams in them. -4.19626 job arises -4.19537 job workers' -3.11480 job good -0.10237 https://github.com/kaldi-asr/kaldi/blob/master/egs/librispeech/s5/local/lm/train_lm.sh

Powered by Blogger