search

thuhcsi / NeuFA

Neural network-based forced alignment with bidirectional attention mechanism
70 stars 8 forks source link

Effect of Reduction Rate #10

Closed ezou626 closed 10 months ago

ezou626 commented 10 months ago

Did you notice any differences in model inference quality with changing the reduction rate hyperparameter of the model? I'm wondering because training takes significantly longer with a small reduction rate.

petronny commented 10 months ago

The reduction rate used here is the same as the reduction rate in Tacotron, which merges some frames of spectrograms into one step.

We use this parameter when we start developing our model to quickly demonstrate our idea.

And any reduction rate greater than 1 will affect the inference quality, since the model will predict the number of frame instead of the actual time for the boundaries. More frames there are, more precise the boundaries will be.

ezou626 commented 10 months ago

Thank you for letting me know! I am currently training with a reduction rate of 4 because I need the model sooner and I'm using it for more of a proof of concept, so less accuracy is okay for now.