🤗 Transformers Wav2Vec2 + PyCTCDecode

Introduction

This repo shows how 🤗 Transformers can be used in combination with kensho-technologies's PyCTCDecode & KenLM ngram as a simple way to boost word error rate (WER).

Included is a file to create an ngram with KenLM as well as a simple evaluation script to compare the results of using Wav2Vec2 with PyCTCDecode + KenLM vs. without using any language model.

Note: The scripts are written to be used on GPU. If you want to use a CPU instead, simply remove all .to("cuda") occurances in eval.py.

Installation

In a first step, one should install KenLM. For Ubuntu, it should be enough to follow the installation steps described here. The installed kenlm folder should be move into this repo for ./create_ngram.py to function correctly. Alternatively, one can also link the lmplz binary file to a lmplz bash command to directly run lmplz instead of ./kenlm/build/bin/lmplz.

Next, some Python dependencies should be installed. Assuming PyTorch is installed, it should be sufficient to run pip install -r requirements.txt.

Run evaluation

Create ngram

In a first step on should create a ngram. E.g. for polish the command would be:

./create_ngram.py --language polish --path_to_ngram polish.arpa

After the language model is created, one should open the file. one should add a </s> The file should have a structure which looks more or less as follows:

\data\        
ngram 1=86586
ngram 2=546387
ngram 3=796581           
ngram 4=843999             
ngram 5=850874              
                                                  
\1-grams:
-5.7532206      <unk>   0
0       <s>     -0.06677356                                                                            
-3.4645514      drugi   -0.2088903
...

Now it is very important also add a </s> token to the n-gram so that it can be correctly loaded. You can simple copy the line:

0 <s> -0.06677356

and change <s> to </s>. When doing this you should also inclease ngram by 1. The new ngram should look as follows:

\data\
ngram 1=86587
ngram 2=546387
ngram 3=796581
ngram 4=843999
ngram 5=850874

\1-grams:
-5.7532206      <unk>   0
0       <s>     -0.06677356
0       </s>     -0.06677356
-3.4645514      drugi   -0.2088903
...

Now the ngram can be correctly used with pyctcdecode

Run eval

Having created the ngram, one can run:

./eval.py --language polish --path_to_ngram polish.arpa

To compare Wav2Vec2 + LM vs. Wav2Vec2 + No LM on polish.

Results

Without tuning any hyperparameters, the following results were obtained:

Comparison of Wav2Vec2 without Language model vs. Wav2Vec2 with `pyctcdecode` + KenLM 5gram.
Fine-tuned Wav2Vec2 models were used and evaluated on MLS datasets.
Take a closer look at `./eval.py` for comparison

==================================================portuguese==================================================
polish - No LM - | WER: 0.3069742867206763 | CER: 0.06054530156286364 | Time: 58.04590034484863
polish - With LM - | WER: 0.2291299753434308 | CER: 0.06211174564528545 | Time: 191.65409898757935

==================================================spanish==================================================
portuguese - No LM - | WER: 0.18208286674132138 | CER: 0.05016682956422096 | Time: 114.61633825302124
portuguese - With LM - | WER: 0.1487761958086706 | CER: 0.04489231909945738 | Time: 429.78511357307434

==================================================polish==================================================
spanish - No LM - | WER: 0.2581272104769545 | CER: 0.0703088156033147 | Time: 147.8634352684021
spanish - With LM - | WER: 0.14927852292116295 | CER: 0.052034208044195916 | Time: 563.0732748508453

It can be seen that the word error rate (WER) is significantly improved when using PyCTCDecode + KenLM. However, the character error rate (CER) does not improve as much or not at all. This is expected since using a language model will make sure that words that are predicted are words that exist in the language's vocabulary. Wav2Vec2 without a LM produces many words that are more or less correct but contain a couple of spelling errors, thus not contributing to a good WER. Those words are likely to be "corrected" by Wav2Vec2 + LM leading to an improved WER. However a Wav2Vec2 already has a good character error rate as its vocabulary is composed of characters meaning that a "word-based" language model doesn't really help in this case.

Overall WER is probably the more important metric though, so it might make a lot of sense to add a LM to Wav2Vec2.

In terms of speed, adding a LM significantly reduces speed. However, the script is not at all optimized for speed so using multi-processing and batched inference would significantly speed up both Wav2Vec2 without LM and with LM.