Montreal Forced Aligner and the state of speech-to-text alignment in 2026

The Montreal Forced Aligner (MFA), first released in 2016, has been updated to version 3.0 with expanded multilingual support and benchmark results that place it at or near the top of forced alignment tools, according to a new paper on arXiv [1]. The paper, submitted on 16 June 2026, documents the tool's evolution over the past decade and evaluates its performance on English, Japanese, and Korean [1]. MFA 3.0 was tested against four benchmark datasets and achieved mean boundary errors below 15 ms, a result the authors describe as state-of-the-art or near state-of-the-art [1]. The system's development since version 1.0 has included the integration of larger open-source datasets, harmonized IPA dictionaries, model adaptation, cross-language phone remapping, and support utilities [2]. The authors report that adaptation and cross-language remapping prove effective for languages outside MFA's training distribution, while pronunciation probability modeling and phonological rules deliver gains in specific conditions [1]. The Montreal Forced Aligner has become the most widely used tool for forced alignment in both research and industry since its initial release [2]. Forced alignment, the process of matching speech audio to its corresponding text transcription at the phone or word level, is a foundational task in phonetics, speech recognition, and linguistic fieldwork. The latest version's performance across typologically distinct languages such as Japanese and Korean suggests broader applicability for under-resourced languages, though the paper's evaluation remains focused on these three languages [1]. The preprint is available on arXiv under the Computation and Language category [1].