Montreal Forced Aligner

Project Overview

This project demonstrates forced alignment of speech audio with text transcriptions using the Montreal Forced Aligner (MFA), including handling of Out-of-Vocabulary (OOV) words.

Dataset

• Audio files: 6 WAV files (3 from F2BJRLP dataset, 3 from ISLE dataset)
• Transcripts: Corresponding text transcriptions
• Total duration: 97.16 seconds

Installation & Setup

Prerequisites

• Windows 10/11
• Miniconda3

Step 1: Install Miniconda

1. Download from https://docs.conda.io/en/latest/miniconda.html
2. Install with default settings
3. Restart terminal

Step 2: Install Montreal Forced Aligner

# Create conda environment
conda create -n aligner -c conda-forge montreal-forced-aligner
# Activate environment
conda activate aligner
# Verify installation
mfa version

Step 3: Download Models

# Download acoustic model
mfa model download acoustic english_us_arpa
# Download dictionary
mfa model download dictionary english_us_arpa
# Download G2P model (for OOV handling)
mfa model download g2p english_us_arpa

Project Structure

MFA-Forced-Alignment-Assignment/
├── audio/                  # Original .wav files
├── transcripts/            # Original .txt transcription files
├── corpus/                 # Paired audio + text files (created during usage)
├── dictionary/             # Custom pronunciation dictionary files
├── output_before_oov/      # TextGrid files (initial alignment)
├── output_after_oov/       # TextGrid files (after OOV handling)
├── screenshots/            # Praat visualization screenshots
├── docs/                   # Additional documentation
├── REPORT.md               # Detailed analysis and observations
└── README.md               # This file

Usage

Step 1: Prepare the Corpus Directory

MFA requires audio and transcription files to share the same base name in a single directory.

mkdir corpus
cp audio/*.wav corpus/
cp transcripts/*.txt corpus/

Step 2: Initial Alignment

mfa align corpus/ english_us_arpa english_us_arpa output_before_oov/ --clean

Step 3: Validate & Identify OOV Words

mfa validate corpus/ english_us_arpa english_us_arpa --ignore_acoustics

This command produces logs including oovs_found.txt with all OOV words and their counts.

Step 4: Handle OOV Words with Custom Dictionary

1. Save the base dictionary locally:

   mfa model save dictionary english_us_arpa dictionary/base_dict.txt

2. Create custom dictionary:

   cp dictionary/base_dict.txt dictionary/custom_dict.txt

   Edit custom_dict.txt and add manual pronunciations for OOV words in ARPABET format.

3. Re-run alignment with custom dictionary:

   mfa align corpus/ dictionary/custom_dict.txt english_us_arpa output_after_oov/ --clean

Results

• Initial alignment: 6 TextGrid files generated in output_before_oov/
• OOV words identified: 22 unique types (48 total tokens)
• Custom dictionary created with phonetic pronunciations for all OOV items
• Final alignment completed with improved TextGrids in output_after_oov/

Visualization

TextGrid files can be opened in Praat for visualization:

1. Download Praat: https://www.fon.hum.uva.nl/praat/
2. Open the corresponding .wav file and .TextGrid file
3. Select both objects and click View & Edit

Screenshots of before/after OOV handling comparisons are stored in the screenshots/ folder.

Key Observations

• Alignment accuracy improved significantly after adding custom pronunciations
• Most OOV issues were caused by proper names and non-standard terms
• Detailed analysis and metrics are provided in REPORT.md

Author

Shashank Vishwakarma

Date

February 2026