ASR Project

ASR Project within a course 'Deep Learning in Audio'

How to recreate the experiment

Crucial information

In ctc_char_encoder.py I added the beam search with and without LM.

What do we need for LM?

pip install https://github.com/kpu/kenlm/archive/master.zip
wget https://www.openslr.org/resources/11/3-gram.arpa.gz --no-check-certificate
gzip -d 3-gram.arpa.gz

How do I download the model?

wget https://drive.google.com/file/d/1FfcDs004kl3bo8prP-TmvpES9igAayBl/view?usp=sharing

How do I run test?

For test-clean:

python3 test.py \
   --config hw_asr/configs/test_clean.json \
   --resume model_best.pth \
   --batch-size 64 \
   --jobs 4 \
   --beam-size 100 \
   --output output_clean.json

For test-other:

python3 test.py \
   --config hw_asr/configs/test_other.json \
   --resume model_best.pth \
   --batch-size 64 \
   --jobs 4 \
   --beam-size 100 \
   --output output_other.json

How do I train the model?

Dependent on which model you want, but I trained the best model using this command

python3 train.py -c hw_asr/configs/train_together.json

How do I evaluate on test data?

python3 test.py \
   -c hw_asr/configs/test_other.json \
   -r model_best.pth \
   -t test_data \
   -o test_result.json \
   -b 5

Device

GPU: Tesla P100-PCIE

Model choice

For my implementation of ASR, I chose the model from paper Deep Speech 2: End-to-End Speech Recognition in English and Mandarin. Model architecture in the project is heavily dependant on this architecture, though a bit reduced due to resources limitations.

Specifications

For padding, kernel sizes and strides in the convolutional layers I looked up the official documentation of DeepSpeech2 in order to build a strong model from the very first steps.

Training pipeline

First method

1. First I am training my model only on LibriSpeech dataset train-clean-100 and train-clean-360 for 50 epochs and testing the model quality on LibriSpeech dataset test-clean. [since the audios are clean i did not add any noise augmentations]
2. Then I am fine-tuning the model on LibriSpeech dataset train-other-500 for 50 epochs and testing its final quality on test-other. [on the fine-tuning step I am adding noise augmentations to upsample the training data and adapt the model to noisy sounds]

Second method

The hypothesis here is that there is a decent shift in the distributions of data when we move from the clean dataset to 'other', so the first method with finetuning + pretraining would not work well. Due to that, there is a need to train on both datasets and make the model train on various examples. So the pipeline is:

1. Train on both clean + other train and eval on test-other straight away to understand the real accuracy of my model while training.

Experiments

First try

First I implemented a model with 2 convolution layers and 4 GRU layers. Optimizer - SGD, lr_scheduler - OneCycleLR, no augmentations for train-clean part. See the config for the clean-part training here. The metrics on this step while training only on clean part were CER: 0.39 and I would like to not state the WER, because I was not computing it in the right way.

Second try

I decided to carry on with training on the clean part and get better cer and wer metrics on them, so I am adding more GRU layers and augmentations. I decided to take the architecture which is shown in the picture in part 1 [3 convolution layers and 7 GRU layers]. First for 50 epochs I trained on the train-clean datasets adding such augmentations: for waves - Shift, Gain and Guassian Noise; for spectrograms: Frequency masking and Time masking. Afterwards, I trained the same model for 50 epochs on train-other with the same augmentations, but without adding the Guassian Noise. Configs for training - first 50 epochs and second 50 epochs. Logs of training: first 50 epochs and second 50 epochs.

Metrics:

[model numbers were first train_ds2/1021_175629, second train_ds2_other/1022_173704]

test-clean

Beamsearch	WER	CER
NO	0.45	0.16

test-other

Beamsearch	WER	CER
NO	0.63	0.27

What failed?

Using TimeInversion Augmentation was a failure, the audio became impossible to listen to and this augmentation was rather vague.

Third try

I decided to try to train the model for 100 epochs on the whole dataset(clean + other) and see how well that goes. Model architecture was a bit changed since i had to downsize the number of GRU layers and make it only 4 due to enlargening the dataset. This attempt rather failed too [see metrics]. logs and config

Metrics

test-clean

Beamsearch	WER	CER
NO	0.99	0.71

test-other

Beamsearch	WER	CER
NO	1.02	0.68

Why failed?

My hypothesis is that using Guassian noise augmentation with the "other" dataset was an overkill. Also, SGD optimizer might be less efficient than Adam in this case, but I wanted to try the CyclirLR scheduler with triangular2 mode and it is compatible only with SGD.

Beam search without LM

You can see the code of my beamsearch without using Language model in the /text_encoder/ctc_char_text_encoder.py file, yet I wouldn't say that custom beam search actually did something valuable here, CER and WER metrics went up with beam search results [example below]

Beam Search [without LM] results example on the model we got from the first experiment

Beamsearch	WER	CER
NO	0.46	0.16
YES	0.47	0.17

[weird, no rise in metrics]

Fourth try

Here we have some hyperparameters games, didn't change much though - only changed the number of epochs + removed the guassian noise augmentation + we are still training on clean + other train, but now in one epoch we have 1000 steps [might be an overkill] + running val on test-other to see how good of a score we can get. Since the training data is large, i removed the number of GRU layers[left 4 GRU layers and 2 convolutional layers] and got left with only 19M params - this might result in insufficient quality.

Metrics

1. Since one epoch lasted for about 20 minutes due to the number of steps per epoch - I decided to stop the training on the 33 epoch to see how well this change in hyperparameters resulted in the model accuracy. On test-clean:

Beamsearch with LM	WER	CER
NO	0.27	0.08
YES	0.16	0.06

This is a solid 7

On test-other:

Beamsearch with LM	WER	CER
NO	0.45	0.19
YES	0.46	0.19

Beam search with LM

Implementation

For beam search with LM i used the 3-gram.arpa LM from LibriSpeech models, implemented a ctc decoder which is able to decode texts. Additionally, I added some hot words[hard words which not a lot of people could guess how to write] which are passed to the decoder with a significant weight for the model to draw attention to these words specifically.

Hyperparameters

In the prior experiment[№4] I implemented Beam Search with LM with hyperparameters alpha = 0.5 and beta = 0.1, maybe this needs some changes so I will experiment with these in the fifth, final experiment.

Best model

Download the model

You can download the model's weights here

Architecture

2 Convolutional layers + 5 GRU layers. In the paper Deep Speech 2: End-to-End Speech Recognition in English and Mandarin authors said that there is no radical gain in 3 convolutional layers and 7 GRU layers, and one can stop on 2 conv layers + 4 gru layers, but I decided to add one more just in case(it didn't slow down the training process too much).

Augmentations

For waves I used Shift, Gain, Guassian Noise[links to their docs are given above] and for spectrograms I used TimeMasking and FreqMasking. Since the dataset i was training on was clean + other, I decided that the probablitity of adding Guassian Noise should not be 1 as in other experiments, because it can make the data too inaudible. You can listen to the audio + look at spectrograms after augmentations here.

Metrics

On test-clean:

Beamsearch with LM	WER	CER
NO	0.22	0.069
YES	0.1996	0.074

On test-other:

Beamsearch with LM	WER	CER
NO	0.43	0.17
YES	0.42	0.198

logs

Config for train, for test-clean and test-other. Training logs are here

encoder

For LM I tried different hyperparameters for alpha and beta, but the best ones were alpha = 0.5 and beta = 0.1, yet i tried different betas from 1e-3 to 0.1

Wandb reports with all the necessary data

for the best run

Here

for all the good runs

Here