And I think this was Applejack's encoder on RJ Banks, giving him an accent.

very interesting! can you post some output from the original models as comparison?

RJ Banks, no pony

Applejack, no Banks

The voice blending was disappointing, it just kinda sounded like the two overlapped, rather than like the throat and mouth of the two averaged out, which I suppose makes sense with it being, what, noise vocoded with spectrograms of voices?

RJ Banks encoder on pony

Accent? Gone

At least that part works well enough so something was gained from fucking with it all.

Seems to be that the encoder* parts impacts accents, and the decoder* parts are the voice timbre

Here's my mess of a notebook, I ain't cleaned it or nothing since it's a sandbox. The blending part checks all the named entries in the model dict and if it doesn't match a name you entered it gets deleted, and what's left is used to replace values in another model. The google drive links to models in the code cell there are my own, except applejack, so don't expect them to stick around forever.

Run

print(torch.load('RJBANKS')['state_dict'].keys())

to get a list of all the parameters in the saved models

decoder.attention_rnn* seems to handle some of the gruffness and emphasis / stress.

decoder.linear_projection* from RJ Banks onto pony, suddenly it's all smoker

I've tried freezing various layers to try to train better on small datasets but it didn't really work well

here's a few samples of my oomerang Annoucer Model with a whopping 86 audio Files

of data

that's John O'Hurley, right? i think he's got some audiobooks that'd work for data

yep

this what i call: perfect

Amazing

Is there any better notebooks or programs than https://colab.research.google.com/drive/18lBRBWOs4uV1DjhoW_fVzoydYUw400PW to transcribe audio?

I realize that The "Josie And the pussycats" had another A in it's Transcript so i fixed It

Now I wait

The mystery of figuring out the transcript problem continues

Would a batch size of 16 be good for 100 wav files?

new training with hand-written transcription

much better

epoch 120

epoch 200 🎉