Skip to content

Releases: spicytigermeat/SOFA-Models

tgm_en_v100

18 Aug 18:23
ed3218d

Choose a tag to compare

New English SOFA model trained on data only labelled by myself or hand verified by myself. I was having lots of issues with past models so I decided to do some experiments. I'll be doing more in the future, but I am happy with the results this got, so I'm gonna come back to it at another point. Trained on SOFA v1.0.3

Utilizes the same dictionary from v005.

Known Issues:

  • Sometimes when an [r] follows a vowel, it doesn't get labelled properly.
  • Needs more high range/femme voice data.
  • Biggest issue is the dictionary. I notice when the phonemes aren't exact, the model struggles to place them, but when they're manually edited in the transcription editor, the placements of phonemes are much more accurate.

Training Data information:

Name | Voice Provider

A. (Unrevealed) | N.
Bitter | Guillotama
Canary | Mina Moonrise
C.B. (Unrevealed) | M.B.
Leif | FerretFather
Luther | imsupposedto
Miyo | ShiWeiMigi
TIGER | tigermeat
TRITON | Ryan M.

v005

10 Jan 03:18
ed3218d

Choose a tag to compare

🎉 tgm_sofa: English SOFA Model by tigermeat, version v005! 🎉

Trained on 11.6 hours of data for ~24hrs, ~1400 epoch/~120,000 steps w/ 1250-1500 max_batch_size on 3090 24gb vram

Dictionary has been SLIGHTLY changed as of now, biggest change being entries for single phonemes, each needing to start with [.], which is helpful for things like glottal stops [q] and vocal fry [vf].
Example (transcription > g2p output): .hh .eh .l .ow world > hh eh l ow w er l d

Data Credits:
NOTE: All data used in this model is used with permission and is ethically collected.

TIGER (tigermeat)
TRITON (Ryan M.)
Canary (MinaMoonrise)
Miyo (shiweimigi)
Leif (Beikon)

Alex Floarea (Alex Floarea)
Fushine Makojo (alice)
TSVD (labels by nobodyP)
Aida (Iris)
Coby (Iris)
Evelyn (Iris)
Moralix (Moralix)

v004

01 Jan 19:48
ed3218d

Choose a tag to compare

🎉 3rd try at this model!

Trained for ~24hrs, 2010 epoch/107,000 steps w/ 2000 max_batch_size on 3090 24gb vram
Dictionary has been unchanged as of now.

Data Credits:
NOTE: All data used in this model is used with permission and is ethically collected.

TIGER (tigermeat)
TRITON (Ryan M.)
Canary (MinaMoonrise)
Miyo (shiweimigi)
Leif (Beikon)

Alex Floarea (Alex Floarea)
Fushine Makojo (alice)
TSVD (labels by nobodyP)

Initial Release

16 Dec 04:29
ed3218d

Choose a tag to compare

First release of this model. 🎉

Strengths:

  • All data is hand verified by me
  • Lots of different singing types, and a few extra phonemes

Weaknesses:

  • Not many speakers
  • Very little femme voice data
  • only about 4 hours of data was used to train this