Skip to content

Latest commit

 

History

History
21 lines (19 loc) · 1.34 KB

MFA.md

File metadata and controls

21 lines (19 loc) · 1.34 KB
Raw

Montreal Forced Aligner: Trainable Text-Speech Alignment Using Kaldi

McAuliffe, Michael, Michaela Socolof, Sarah Mihuc, Michael Wagner, and Morgan Sonderegger.

In Interspeech, vol. 2017, pp. 498-502. 2017. [PDF]

Whats Unique This paper present the tool to align orthographic transcript at word level and phonene level to the speech. And, it retains trainability with new datasets.

Key Take Aways

  • MFA has been scuccessfully applied to 29 languages.
  • Its training process involve following steps:
    • 40 iterations of monophone GMM training
    • 15 iterations of monophone realignment
    • 35 iterations of triphone training
    • 15 iterations of triphone realignment
    • 35 iterations of speaker-adapted triphone training
    • 15 iterations of speaker-adapted realignment
  • It uses mel-frequency cepstral coefficients (MFCCs) as acoustic features.
  • Training on LibriSpeech corpus of 1000 hours data took 80 hours.
  • It is evaluated against the manual annotation, Buckeye Corpus contains 20.7 hours of conversational speech from 40 speakers with manual transcript and boundaries at phone and word level.
  • Average difference in actual and word level boundary is around 25 ms, and for phoneme it is around 17 ms.
  • It uses Kaldi toolkit for underlying audio processing.