A female vocal LoRA named "Lina".
This LoRA is a versatile female voice model capable of "singing" in three languages: Lithuanian (I specifically trained it for this), Russian, and English.
It sings in Lithuanian and Russian with absolutely no accent; I cannot say the same with 100% certainty for English, as I do not speak the language perfectly myself.
The LoRA was trained on a mix of digitized analog (50%) and digital (50%) audio tracks, giving it characteristic "analog" sound qualities (a deliberate training choice).
The LoRA trigger word is linos_vokalas. It was trained in 2,500 steps.
Suggested prompt keywords: velvet and smoky female voice, very soft and breathy vocal texture, bright and resonant clear singing, powerful dramatic vocal depth, clean acapella, professional studio acoustics.
This is a strong LoRA; a weight of 1.0 might introduce audio artifacts. It works well with other LoRAs, though using it alongside LoRAs optimized for other languages might result in a noticeable "accent" in the vocals.
P.S. This is my first LoRA, so I would greatly appreciate your feedback, comments, and advice.
Description
"linos_vokalas_2" LoRA.
This is a synthetic female vocal, generated from 5 different vocal styles (ranging from 'velvet' to 'soft'). Some of the source vocals were taken from digitized audio tape recordings to achieve an analog vocal sound.
Trained for 4,800 steps at a rate of 0.00005.
The LoRA is bilingual—it 'sings' well in Lithuanian and Russian. I cannot say whether it sings in English without an accent, as English is not my strong suit.
Trigger word: linos2_vokalas.
Required prompt keywords: velvet and smoky female voice, very soft and breathy vocal texture, bright and resonant clear singing, powerful dramatic vocal depth, clean acapella, professional studio acoustics.
Comments (3)
This is a LoRa for the YuE2 music model.
For some reason, I couldn't publish the description for linos_vokalas_2 properly, so I'm posting it here in the comments—the trigger is linos2_vokalas.
This version is weaker, even though it was trained using a higher number of steps.
The first version was "overlearned," so it could sound muddy with certain instruments (like guitars in rock or metal music).
I didn't notice that "overlearning" issue in the second version linos_vokalas_2).
If you have any feedback, please leave a comment—I'll try retraining it.
Intermediate in strength version between Linos_vokalas and Linos_vokalas_2 — Linos_vokalas_3.
0.0005 x 1400 steps.
Trigger: linos3_vocal
0.8–1 is necessary for Lithuanian pronunciation. For lighter variants — between 0.4 and 0.8.
If you use it alongside Mindaugas’s vocals at equal levels, it will overpower the male vocal and ‘sing’ (from what I’ve tried) with an incredibly ‘velvety’ voice—almost masculine in tone.