Datasets:

rhasspy

/

piper-checkpoints

like 40

Follow

Rhasspy 35

I have a checkpoint, but it is for the "x-low" audio quality and has fewer phonemes than the current Piper models. I think you will actually have better luck starting from one of the English Lessac checkpoints and fine-tuning Italian.

Hi Michael,

I’m reaching out to ask for some support with Piper TTS. I’m not experiencing any particular issues, but I’d like to ensure I’m getting the best possible results when cloning a voice.

I’ve done all my recordings using piper-recording-studio in Italian, and I have 1150 recordings in total. I’ve noticed that the sample rate of the recordings is 48000 Hz, so I’m wondering if setting -sample_rate 22050 during piper_train.preprocess will automatically normalize the recordings to the correct sample rate.

Also, would it make sense to enable validation to monitor for potential overfitting?

At the moment, I’ve trained on 500 recordings with a batch size of 22, and the loss values I’m sharing in the attached image reflect about 3000 epochs of training. I’m doing fine-tuning on a medium English model, but I previously tried the high model, where I achieved slightly better results in fewer epochs (about 1000) with a batch size of 16.

Lastly, I wanted to ask about the dataset size needed to train a model from scratch. I’m planning to create an Italian model from scratch, as I’ve found a corpus of 59,000 multi-speaker recordings.

Thank you so much for your support!

Hi MicheleNas,
I'm very interested in using a medium-to-high quality Italian voice for Piper TTS. While I don't have experience or dedicated hardware for training voices, I would love to contribute to your project in any way I can. Whether it's through testing, providing feedback, or supporting other aspects of development, I'm eager to help bring a high-quality Italian voice to life.
Please let me know if there's a way I can assist!

Your training graphs look good to me! I think using 22050 Hz would probably train faster, and may let you fit a larger batch size on your GPU.
In general, I find that having more clean and consistent samples is the best way to improve the voice. Clean both in the recording and in the transcripts. And consistent in the sound of the room, how close you are to the microphone, etc.
If you're interested in donating your voice, I'd be happy to train a model for you as well on my GPUs :)

Hi @synesthesiam ,
I, like @MicheleNas , am also trying fine-tuning on an Italian dataset, but using lessac-high checkpoints as a basis.
In particular, I have selected a single voice from the CML-Multi-Lingual-TTS, which seems particularly good to me.
I am fine-tuning on a GPU that is not particularly powerful, so I think it will take a few days to get a first result, however the curve of the loss function looks good.
If I can, I will also select some other quality voices in the near future, so that I can expand the choice for the Italian voice.