First get a sense for what Cookie's script does with the files. Run the first 4 cells in Cookie's script once normally using all the defaults. Notice the lines: >with open('/content/tacotron2/filelists/clipper_train_filelist.txt', 'w') as train_out: > train_out.write(';\n'.join(train)+';\n'.join(train_arpa)+";") >with open('/content/tacotron2/filelists/clipper_val_filelist.txt', 'w') as val_out: > val_out.write(';\n'.join(validation)+';\n'.join(validation_arpa)+";") These files tell tacotron where the clips are and what their corresponding transcripts are. If you open each of these files, you'll see a long list of lines like this: >wavs/Celestia-fim:s8e7-896.778363-899.650633.wav|Twilight, you felt what i was emoting, didn't you?; In this case, the audio clip is in /content/tacotron2/wavs/Celestia-fim:s8e7-896.778363-899.650633.wav, and the transcript is "Twilight, you felt what i was emoting, didn't you?" So you'll want to put all your audio clips in /content/tacotron2/wavs/, and you'll want to generate files with one line per clip in this format: >wav/clip-name.wav|transcript; >wav/clip-name2.wav|transcript2; >... You'll want to create two such files. One of the files should contain roughly 4157 lines (95% of the clips), and the second should contain roughly 219 lines (5% of the clips). The larger one should be written to /content/tacotron2/filelists/gothic_train_filelist.txt, and the smaller one to /content/tacotron2/filelists/gothic_val_filelist.txt. Further down, you'll see >hparams.training_files = "filelists/clipper_train_filelist.txt" >hparams.validation_files = "filelists/clipper_val_filelist.txt" Replace "clipper" with "gothic" in both lines. Then let it run.