Skip to content

Conversation

lottev1991
Copy link
Contributor

@lottev1991 lottev1991 commented Oct 1, 2025

Changes

  • New Italian G2P:
    • The G2P has been re-trained from scratch (50 batch size, 300 epochs) and should give improved results.
    • It was based on the same dictionary as before, but it has been edited:
      • The vowels with stress markings have been merged with their unmarked versions, as in the Arpabet G2Ps.
      • The rr phoneme has been removed, since it does not actually exist in Italian (it was absent in the original CMUdict, so this was essentially a revert). The geminated r is treated as a sequence of two of the same phoneme instead.
      • The y phoneme has been reverted to j, as in the original CMUdict.
      • Affricates (dz, ts, dZZ, tSS), when geminated (doubled), now have a plain stop consonant before them (e.g. [d dz], [t ts], etc.). This is more linguistically accurate, should give no issues, and works better with DiffSinger phoneme merging (since affricate doubling rules can vary per language).
  • The phonemizers that are reliant on this G2P have been updated accordingly, so they should continue to work as intended.
  • I also added some CC fallbacks to the Italian SBP.

Note

The allophones nf and ng are still present, but they mostly exist for concatenative synthesis. They can be easily merged with n for DiffSinger without problems, if proper pronunciation rules are followed.

@lottev1991 lottev1991 marked this pull request as draft October 1, 2025 17:01
@lottev1991 lottev1991 marked this pull request as ready for review October 1, 2025 17:39
@lottev1991 lottev1991 marked this pull request as draft October 1, 2025 18:41
@lottev1991 lottev1991 marked this pull request as ready for review October 1, 2025 18:53
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant