Uberduck

each line contains a file path and the transcription

so to be clear, that txt file that prepare mels expects has the audio file name and transcribed text pipe separated

yeah exactly. and that's what transcribe_dataset.py will output

Using GPT3 to generate jeopardy style questions and answers based on topics that don't actually appear in jeopardy...

```
Q: Gradient descent can be expressed in terms of this quantity.
A: What is the loss function?
R: No, the correct answer is the hessian matrix.

Q: An artificial neural network that predicts the next symbol in a sequence based on the previous symbols is called this.
A: What is a recurrent neural network?
R: That was the correct answer.
```

questions, answers and followup remark were all generated by GPT3

Definitely need to work on improving pronunciation but pretty promising start

to avoid any copyright issues we can call it GPrT (pronounced "jeopartee")

it isn't public, but they keep expanding the private beta

I think there's probably around 10k people in it by now

don't think they'll release the model itself though

ah, I see. I thought they stopped when I saw the microsoft announcement

I started training trebek on a dataset with arpabet transcriptions to try to improve pronunciations

fairly mediocre results so far but I think that's because I don't have enough data (~1 hour or so), and the base LJSpeech model wasn't trained on arpabet at all. At least it's learning some of the tokens

```
  "{AY1}, {Z AE1 K}, {W AH1 N S} {W AA1 Z} {AH0} {M AE1 N}... {DH EH1 N} - {L EY1 T ER0}, {AY1} {B IH0 K EY1 M} {B IY1 S T}!!"
  "{M AY1} {F EY1 V ER0 IH0 T} {W ER1 D} {IH1 Z} {AE2 N T AY0 D IH2 S AH0 S T AE2 B L IH0 SH M AH0 N T EH1 R IY0 AH0 N IH2 Z AH0 M}."
  "{M AY1} {F EY1 V ER0 IH0 T} {HH AH0 W AY1 AH0 N} {AY1 L AH0 N D} {IH1 Z} {M AW1 IY0}."
```

I decided to try to improve on this approach by training on LJSpeech with both normal and arpabet transcriptions .... here's a sample from 5 epochs in 😁

never trained tacotron from scratch before so this'll be interesting! hopefully she'll learn to speak