Montreal Forced Aligner 3.0 achieves state-of-the-art with sub-15ms errors
A decade-old tool just got a major upgrade — now supporting cross-language alignment with record accuracy.
The Montreal Forced Aligner (MFA), originally released in 2016, has become the standard tool for forced alignment in speech research and industry. In its latest version — MFA 3.0 — the tool has undergone substantial development, expanding language coverage and improving accuracy. A new paper documents these changes and benchmarks the system against classic and neural forced aligners across English, Japanese, and Korean.
MFA 3.0 achieves state-of-the-art or near state-of-the-art performance on all four benchmark datasets, with mean boundary errors below 15 milliseconds. The system now features harmonized IPA dictionaries, model adaptation for new languages, and cross-language phone remapping. Pronunciation probability modeling and phonological rules provide additional gains under specific conditions. These updates make MFA 3.0 significantly more robust and useful for multilingual speech-processing tasks, without sacrificing the low latency that made its predecessors popular.
- Mean boundary errors below 15 ms on all four benchmark datasets (English, Japanese, Korean).
- Expanded coverage via larger open-source datasets, harmonized IPA dictionaries, and cross-language phone remapping.
- Model adaptation and pronunciation probability modeling enable strong performance on languages outside MFA's original training distribution.
Why It Matters
MFA 3.0 sets a new accuracy baseline for forced alignment, enabling better speech analysis and ASR across languages.