@hecko do you use ARPA?

Or MFA in any way

Also did you take out all MLP clips less than 2 seconds?

i haven't figured out how to use arpabet with tacotron yet forced aligner idk what i'd need for and i did not, in part because most of them are short anyway and also because i didn't think of it

i could add #3 to the notebook though

Oh my god, I remember that voice.

hello @User , I have a question about this graph

looks good to me

idk why you asked me specifically though

Haahaa coz you helped me out last time

Anyway so my question is , where is it from?

Is it from the evaluation data, test data or train data?

(sorry for not responding earlier) i think it's from validation data, but that doesn't really matter because the notebook uses training data as validation data anyway

oh ok thanks . maybe my question would be so what are they actually plotting. Is it the encoder and decoder losses?

the x axis is time and the y axis is attention, essentially what characters the network is looking at while reading

so the horizontal moments are likely when it encounters a comma or period

oh sounds confusing, thanks for breaking it down

initially I thought since the labels have encoder vs decoder timestep and the encoder time steps seem less than those of the decoder, then probably it moves less steps than the encoder

just double checked and it seems to be like

the text is input into the encoder which pre-chews the data and turns it into phoneme-like things and such

and then once it's encoded the decoder looks at that at its own pace

so yeah the captions are confusing

personally i prefer the talknet approach, where there's one bit that predicts the duration of each phoneme, then they're all copied and pasted enough times to meet the predicted duration, so the spectrogram generator doesn't need to learn to read in order it can just convert each bit of the input into a bit of the spectrogram (i am bad at explaining) but admittedly that requires an external mechanism to figure out the durations of the training audio's phonemes to train the duration predictor, whereas tacotron does it by itself

isnt talknet meant for producing music like sound??

Thought tacotron was built for more natural voice and talknet for music

both is wrong^^

gee i wonder what a nn named *talk*net could be good for

the singing capabilities are just a fun side effect of the segmentation into duration and pitch and spectrogram networks