BUG REPORT: Generates MALE vocal when FEMALE selected.

[merket]

This happens a lot, I am not sure if it's regarding the training data bias but even though the vocal is selected as "female" and parsed correctly to the terminal, the model generates male vocals. Even if "female" is added to the custom styles and parsed to the terminal a second time ( "descriptions": "female, Female,) it can still produce a male vocal.

I don't know if it's the same situation with the male vocals (male selection produces female). Can anyone confirm that and if this is a training bias or a bug or not?

[zhangsh1416]

I'm also struggling with it, it seems that it happens only for female, when you select male, it's always male voice.

[merket]

I'm also struggling with it, it seems that it happens only for female, when you select male, it's always male voice.

Since I wrote this, I tried 6 more new generations with the same settings;
"genre": "Rock",
"gender": "female",
"timbre": "Breathy",
"emotion": "Romantic, Tender",
"bpm": 102,
"instruments": "Electric Guitar, Drums",
"output_mode": "separate",
"num_sections": 11,
"has_lyrics": true,
"total_lyrics_length": 1240,
"used_model_server": true,
"reference_audio_id": null

ALL OF THEM ARE MALE!
If I go back to my theory of it possibly being training data bias, I can't wrap my head around how romantic rock songs wouldn't have any female input... 6 generations, I'm telling you... It is really disheartening to wait for all of them to hear it's not at all what you wanted to generate. The main, core setting of the song; vocal, is delivered opposite... Go figure.

[zhangsh1416]

I've tried to use female + genre + instrument as description(avoid timbre and emotion), things got better, previously 5/7 failed, after changing, 2/7 failed.
Same thing happened to Suno as well: https://www.reddit.com/r/SunoAI/comments/1n5f2i0/i_put_female_vocals_in_the_song_description_but_i/
It was from several months ago, they probably solved this problem now.