How to train multi speaker datasets?

[tuna2134]

No description provided.

[shivammehta25]

Hello,

Check the dataloading:

Matcha-TTS/matcha/data/text_mel_datamodule.py

Lines 166 to 170 in 7780426

	filepath, spk, text = (
	filepath_and_text[0],
	int(filepath_and_text[1]),
	filepath_and_text[2],
	)

The dataset's file list should be arranged in rows like:

text: str|spkd_id: int|audio_location: str 

So example:

Hello, I am Shivam|0|shivam_speech.wav
Hello, I am not Shivam|1|not_shivam_speech.wav
Hello, I am Batman|2|batman_speech.wav
I am not scared of bats now|2|batman_speech2.wav

Update the fielist here:

Matcha-TTS/configs/data/vctk.yaml

Lines 7 to 8 in 7780426

	train_filelist_path: data/filelists/vctk_audio_sid_text_train_filelist.txt
	valid_filelist_path: data/filelists/vctk_audio_sid_text_val_filelist.txt

set the number of speakers accordingly like in the example it is 3 speakers: [0, 1, 2]

Matcha-TTS/configs/data/vctk.yaml

Line 11 in 7780426

n_spks: 109

Then follow the steps in readme to Train with your own dataset. Hope this helps :)

[tuna2134]

Hello,

Check the dataloading:

Matcha-TTS/matcha/data/text_mel_datamodule.py

Lines 166 to 170 in 7780426

	filepath, spk, text = (
	filepath_and_text[0],
	int(filepath_and_text[1]),
	filepath_and_text[2],
	)

The dataset's file list should be arranged in rows like:

text: str|spkd_id: int|audio_location: str 

So example:

Hello, I am Shivam|0|shivam_speech.wav
Hello, I am not Shivam|1|not_shivam_speech.wav
Hello, I am Batman|2|batman_speech.wav
I am not scared of bats now|2|batman_speech2.wav

Update the fielist here:

Matcha-TTS/configs/data/vctk.yaml

Lines 7 to 8 in 7780426

	train_filelist_path: data/filelists/vctk_audio_sid_text_train_filelist.txt
	valid_filelist_path: data/filelists/vctk_audio_sid_text_val_filelist.txt

set the number of speakers accordingly like in the example it is 3 speakers: [0, 1, 2]

Matcha-TTS/configs/data/vctk.yaml

Line 11 in 7780426

n_spks: 109

Then follow the steps in readme to Train with your own dataset. Hope this helps :)

Thanks