Forced Alignment

Lets see next process ,download latest Montreal Force Aligner source code from

Link :- https://github.com/MontrealCorpusTools/Montreal-Forced-Aligner.git

Pre-trained model I used is Spanish you can download it from MFA website.

before doing this check whether you have latest python module python 3.x

if not installed that first using sudo apt-get install python3 .

Also you should have pip package above pip3.x.

pip install --upgrade pip

1. Open terminal and change directory to unzipped folder of MFA.

2.Run bin/mfa_align .

Such usage information should be printed out.

If not change directory and follow third step

3.

cd/path/to/Montreal-Forcced-Aligner/thirdparty

Then Run

python3 kaldibinaries.py ~/kaldi/root

Now change directory to Montreal-Forced-Aligner and run following command.

It would download all packages mentioned inside requirements.txt.

Now run freezing/freeze.sh there will be folder created by the name dist change directory to that folder and further change directory to folder named montreal-forced-aligner and change directory to bin this should contain two files mfa_align and mfa_train_and_align .

Now run bin/mfa_align or bin/mfa_train_and_align from montreal-forced-aligner directory output should be similar to output message mentioned in step 1.

With all steps done under Kaldi compile and Montreal Aligner setup now is the time to align

any given .wav file and .lab file (which contains transcript of given wav file).

NOTE :>

The wav file mentioned should be atleast 16kHz sampling rate.The aligner ignores the files

having sampling rate below 16kHz. The features that the aligner extracts are MFCCs dervied from

spectra from 0 to 7.8 kHz, so sound files must have Nyquist frequencies higher than 7.8 kHz.

To find a sample rate in a Linux system : Right click the file, then click "Properties." The sample

rate information may appear in a different tab of the Properties box depending on your Linux

distribution and version. If it doesn't appear in the main tab, look for an "Audio" or "More Info"

tab.

Let’s start aligning.

My Home directory has

FOLDER CONTENTS

Montreal-Forced-Aligner-master source code from github.

MFATestWav containing .wav and .lab file in spanish

(corpus path)

SpanishDict contains spanish dictionary in text format

(dictionary path).

SpanishModel containing spanish.zip(final.mdl ,

final.occs,meta.yaml,tree)

(pretrained model downloded directly

from MFA website).

(acoustic model path).

First Mono-Phone Training.

Second Tri-Phone Training.

Third Speaker-Adapted Tri-Phone Training.

Command is :>

bin/mfa_train_and_align -t ~/output -v ~/path/to/corpus ~/path/to/dict.txt output_directory

You can also use other command arguments such as

-s > use to allot number of characters to use to identify speakers.

-t > Temporary directory to use for aligining .(~/Documents/MFA).

-j > Number of jobs to use. (default = 3).

-v > verbose.

-c > Temporary files in ~/Documents/MFA and the output directory will be removed prior

to aligning. This is good to use when aligning a new dataset, but it shares a name

with a previously aligned dataset.

-h > Display help message.

-i > ignore.

-q > quiet.

USING PRETRAINED MODEL:

bin/mfa_align corpus_directory dictionary_path acoustic_model_path output_directory

output.zip contains output acoustic model using mfa_train_and_align .Similarly we can use pre-trained acoustic model.

Now to check output , in Montreal-Forced-Aligner-master under folder named spanish/MFATestWav/1.TextGrid

Output of MFA is always in textgrid form ,to check alignment results we use praat.

Praat is used for Speech Analysis,Speech Synthesis,Speech Manipulation,Labelling and Segmentation,etc.

Our purpose is related to analysis.Lets view output of textgrid.

Run :