Loss doesnt goes down

[FacundoMartinezCampos]

I am finetuning a catalan checkpoint from MMS, with a single speaker dataset of 5 hours, the loss doesnt goes bellow 22, all the characters of my the transcription from my datset exist inside the vocab of the model.

I dont know if i just need more training, how many epochs for 2500 audios is a good amount to get better results than the checkpoint?

I already did 400

[omerarshad]

were you able to solve this? I am also finetuning the model on 5k audios, and loss is stuck at 32

[muhammadsaadgondal]

Did you guys make it lesser? I got the same problem
My generated audio sounds a little distorted. Is this also because of it. I trained on 100 epochs and 2e-4 lr

[omerarshad]

Reducing the weight co-efficients works, but you need to train for a lot more epochs then mentioned, and depends on quality of your dataset.

[muhammadsaadgondal]

@omerarshad I'm training on 40 mins of good quality dataset how many epochs do I need to get clear audio.
Can you give the updated weights? I cant locate the saved weights in my local files.

[andergisomon]

@omerarshad I'm training on 40 mins of good quality dataset how many epochs do I need to get clear audio.
Can you give the updated weights? I cant locate the saved weights in my local files.

How long did it take you and what was your hardware setup? I'm wondering if this is possible on a midrange gaming laptop.

[omerarshad]

@andergisomon I finetuned it on google colab pro, A100 with batch size of 64 for around 400+ epochs

[FacundoMartinezCampos]

lowering the mel weight from 35 to 10 after training 2000 epochs on 1 hour of audio made the step loss go from 16 to 7, i need to finish training to see the results but looks promising

[FacundoMartinezCampos]

nope, even when the lossrate dropped from 16 to 7 the quality is still bad, very robotic

[FacundoMartinezCampos]

Reducing the weight co-efficients works, but you need to train for a lot more epochs then mentioned, and depends on quality of your dataset.

which weights?