1 / 18

Presenter : Pei- ning Chen NTNU CSIE SLP Lab

Error approximation and minimum phone error acoustic model estimation Matthew Gibson and Thomas Hain. Audio, Speech, and Language Processing, IEEE Transactions . Presenter : Pei- ning Chen NTNU CSIE SLP Lab. Outline . Introduction Minimum Phone Error Theory Error Approximation

elma
Télécharger la présentation

Presenter : Pei- ning Chen NTNU CSIE SLP Lab

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Error approximation and minimum phone erroracoustic model estimationMatthew Gibson and Thomas Hain Audio, Speech, and Language Processing, IEEE Transactions Presenter : Pei-ning Chen NTNU CSIE SLP Lab

  2. Outline • Introduction • Minimum Phone Error Theory • Error Approximation • Limitation of Baseline Approximation Error • Alternative Error Approximations • Experiments • Error Approximation Analysis • Summary and Future Work

  3. Introduction • Acoustic models estimated using the MPE technique have displayed significant classification performance improvements over ML-estimated models. • This paper introduces a novel error approximation method and demonstrates how it addresses limitations of a previously used technique, and the method is found to yield significant performance improvements when deployed for MPE acoustic model estimation.

  4. MPE • The MPE criterion • : Levenshtein distance

  5. Error Approximation • Alignment-based error approximation:

  6. A substitution example:

  7. Swap the reference and the hypothesis:

  8. A insertion example:

  9. A deletion example:

  10. Limitations of baseline

  11. Frame Error Normalisation • With deletion

  12. With insertion

  13. Using Multiple Reference Alignments • MSNFR and AMSNFR

  14. Analysis • S : substitution, I : insertion, D : deletion

  15. Reference with silence

  16. Evaluation results • Unsmoothed

  17. I-smoothing

  18. Summary and Future work • Significant improvements over the previously introduced error approximation when the symmetrically normalised frame error approximation is deployed for MPE acoustic parameter re-estimation. • Future work should compare use of the approximate methods introduced in this paper with lattice manipulation approaches and the minimum phone frame error.

More Related