Forced Alignment using Montreal Forced Aligner (MFA)

Objective

The goal is to perform forced alignment of speech audio with its corresponding text transcript using Montreal Forced Aligner (MFA).
Forced alignment automatically determines the start and end times of words and phonemes in an audio recording, given the transcript.

---

Dataset

• Audio files (.wav) and corresponding transcripts (.txt) are provided.
• Number of audio files: 6
• Number of speakers: 1
• Each transcript corresponds to a single utterance.

---

Folder Structure

forced_alignment_mfa/
├─ dataset/
├─ outputs/                
├─ praat_screenshots/
├─ scripts/
|  ├─ automate_alignment.py
|  └─ run_alignment.bat
└─ README.md

---

Setup Instructions

1. Install Conda (if not already installed)

Miniconda installation guide

2. Create Conda Environment and Install MFA

conda config --add channels conda-forge
conda create -n aligner montreal-forced-aligner
conda activate aligner

3. Download Pretrained Models

mfa model download acoustic english_us_arpa
mfa model download dictionary english_us_arpa

---

Running Forced Alignment

You have two options to run the alignment: manually using the MFA CLI command or automatically using the provided Python and batch scripts.

---

Option 1: Run Forced Alignment (Manual)

mfa align --clean --overwrite "dataset" "path_to_dictionary/english_us_arpa.dict" "path_to_acoustic_model/english_us_arpa.zip" "outputs"

• dataset: Path to the folder containing audio and transcript files.
• outputs: Folder where TextGrid alignment files will be saved.

---

Option 2: Automation (Extra Credit Work)

To improve efficiency and reproducibility, the entire alignment process was automated using a Python script and a Windows batch file.

Python Script - automate_alignment.py

This script dynamically detects the dataset and output paths relative to its location, making it fully portable across systems.

This script runs the complete MFA pipeline with the --clean and --overwrite flags for all .wav and .txt pairs in the dataset folder.

Click here to view the python script

import os
import subprocess

# Get the directory where this script is located
base_dir = os.path.dirname(os.path.abspath(__file__))

# Move one level up (to the project root)
project_root = os.path.dirname(base_dir)

# Build dynamic paths
dataset_path = os.path.join(project_root, "dataset")
output_path = os.path.join(project_root, "outputs")

# Expand user path for pretrained models (cross-platform)
dict_path = os.path.expanduser(r"~/Documents/MFA/pretrained_models/dictionary/english_us_arpa.dict")
acoustic_path = os.path.expanduser(r"~/Documents/MFA/pretrained_models/acoustic/english_us_arpa.zip")

# Build the MFA command
command = [
    "mfa", "align", "--clean", "--overwrite",
    dataset_path, dict_path, acoustic_path, output_path
]

print("Running MFA forced alignment...")
print("Dataset:", dataset_path)
print("Dictionary:", dict_path)
print("Acoustic Model:", acoustic_path)
print("Output:", output_path)

# Run the command
subprocess.run(command, check=True)
print("\n✅ Alignment completed successfully!")

Batch File - run_alignment.bat

This batch file activates the Conda environment and runs the Python automation script automatically.

Click here to view the batch file

@echo off
call conda activate aligner
python automate_alignment.py

How To Use

Simply double-click the run_alignment.bat file to automatically:

• Activate the Conda environment
• Run the alignment script
• Generate aligned .TextGrid files inside the outputs/ folder

---

Inspecting Results

• Generated TextGrid files can be viewed using Praat software.
• Screenshots of alignments in Praat are included in the praat_screenshots/ folder.

Sample Alignment (ISLE_SESS0131_BLOCKD02_01_sprt1.TextGrid):

Words:

0.0 – 0.44     (silence)
0.44 – 0.53    i
0.53 – 0.92    said
0.92 – 1.33    white
1.33 – 1.48    (silence)
1.48 – 1.8     not
1.8 – 2.25     bait
2.25 – 4.125   (silence)

Phonemes:

0.0 – 0.44     (silence)
0.44 – 0.53    AY1
0.53 – 0.71    S
0.71 – 0.79    EH1
0.79 – 0.92    D
0.92 – 1.03    W
1.03 – 1.17    AY1
1.17 – 1.33    T
1.33 – 1.48    (silence)
1.48 – 1.55    N
1.55 – 1.62    AA1
1.62 – 1.8     T
1.8 – 1.84     B
1.84 – 2.05    EY1
2.05 – 2.25    T
2.25 – 4.125   (silence)

---

Observations

• MFA successfully aligned all audio files.
• The automation script ensures consistent, repeatable results and reduces manual effort.
• Dynamic path detection allows this script to run seamlessly on any system.
• Viewing results in Praat provided a clear visual representation of word and phoneme boundaries.
• Minor warnings (due to single-speaker data) did not impact alignment quality.

---

References

• Montreal Forced Aligner Documentation: https://montreal-forced-aligner.readthedocs.io
• Praat Software: http://www.fon.hum.uva.nl/praat/