Problems training a model with Common Voice

TTS (Text-to-Speech)
#1

Hello,
I want to use Mozilla TTS with Mozilla Common Voice (German).
When I start the training I get the following error: Pastebin: error log. My config.json can be found here: Pastebin: config.json.
Thank’s for your help.

#2

Check if that’s a valid path for the wav file. This usually happens because of wrong paths or corrupt files. If it’s a valid file path, remove it from the train or test csv

#3

Hi, thank’s for your answer.
I already removed one file from the train and validate.csv, but then the same error happens again with another file. I think now that the problem is that the dataset by Mozilla only contains mp3 files.
Greetings

#4

If that’s happening always, you chance that path isn’t valid. If that’s only happening for a couple, then remove them from the csv.

#5

It’s happening always. The path to the folder which contains the audio files is correct. I think the problem is that Mozilla TTS tries to open .wav files, but the dataset only contains .mp3 files.

#6

I looked at the error again. That path is absolutely invalid. Its looking for a ‘.*.mp3.wav’ which it’ll not find there. Go to TTS/datasets/preprocess.py and in common_voice function, make sure the wav file variable will have a proper path, I.e., path/to/mp3.

#7

Thank you for your answer.
I changed the following line in preprocess.py
wav_file = os.path.join(root_path, "clips", cols[1] + ".wav")
to
wav_file = os.path.join(root_path, "clips", cols[1] + "")
and then the path is correct. But now I get the following error:
RuntimeError: Error opening 'CommonVoice/clips/common_voice_de_18482210.mp3': File contains data in an unknown format.
Greetings

#8

Can I have the stack traceback.

#9

Also, I believe it’s because you’re trying to open mp3 with soundfile. Anyway, go to utils/audio.py and try opening with librosa. Should work.

#10

I have made the changes you recommend. But now I am getting the following error:
error log. My audio.py can be found here: audio.py.
Thank you!

#11

Change sample rate in config to 48k instead of 22050.

#13

Hi Timo,
I have the same error while trying to train with the italian common voice dataset. I implemented your same solution but I still have the same error. Did you modify anything else inside the common_voice function in preprocess.py?
Thanks

#14

Hello,
I don’t think so. But I am not sure, it is long ago.
Greetings,
Timo