Word-level timestamps from WhisperX are inaccurate compared to Montreal Forced Aligner (MFA)

Hi,

I’ve been experimenting with Montreal Forced Aligner (MFA) and WhisperX on the same .wav audio file to compare word-level timestamp accuracy.

While MFA produces highly accurate word-level timestamps, I’ve noticed that the timestamps generated by WhisperX are significantly off. I’ve attached a comparison table (see image) showing the differences in word boundaries between MFA and WhisperX for the same audio.

My WhisperX command:

whisperx source.wav --model large-v3 --language en --align_model WAV2VEC2_ASR_LARGE_LV60K_960H --highlight_words True

Questions:

• Given forced alignment is an intended use case for WhisperX, am I missing any recommended settings or workflow to improve alignment accuracy?

I was expecting WhisperX to generate identical word-level timestamps. I have used Praat to visualize the MFA output and found it to be very accurate.

Any guidance or clarification would be greatly appreciated!

Thank you.