Installation on Ubuntu 24.04 LTS

[crispy78]

I'm currently installing Piper on Ubuntu 24.04 LTS and I struggle with the documentation (https://github.com/rhasspy/piper?tab=readme-ov-file#running-in-python). Thankfully there is Thorsten-Voice with his video (https://www.youtube.com/watch?v=b_we_jma220) explaining how to install it. Unfortunately he's running Ubuntu 20.x and I'm running 24.x. These are the steps I had to make in order to get to the same point as Thorsten when he's training the model (at 10:43).

sudo add-apt-repository ppa:deadsnakes/ppa
sudo apt install python3.10
sudo apt install python3.10-dev
sudo apt install virtualenv
wget https://github.com/rhasspy/piper.git
cd piper/src/python
virtualenv -p python3.10 ~/piper/src/python/.venv
source .venv/bin/activate
pip3 install --upgrade pip
pip3 install --upgrade wheel setuptools
pip3 install pip==24.0
pip3 install -r requirements.txt
cd ~/piper/src/python
./build_monotonic_align.sh
Ensure you have espeak-ng installed
sudo apt-get install espeak-ng
pip3 install -e .

mkdir ~/ThorstenVoice-Dataset
wget -O ~/ThorstenVoice-Dataset/thorsten-neutral_v03.tgz https://zenodo.org/records/5525342/files/thorsten-neutral_v03.tgz?download=1
cd ~/ThorstenVoice-Dataset
tar zxvf thorsten-neutral_v03.tgz

python3.10 -m piper_train.preprocess --language de --input-dir ~/ThorstenVoice-Dataset/thorsten-de_v03/ --output-dir ~/piper/src/python/out-train --dataset-format ljspeech --single-speaker --sample-rate 22050

You can pauze the process by pressing CTRL + Z, then run bg to continue this process in the background. Another option is to add " &" (without quotes) at the and of the command to start it in the background immediately. This can be usefull if your SSH-session will time out, laptop is running out of battery, etc. The process takes about an hour running on a Ubuntu 24.04 LTS VM (4 CPU's, 8192 GB memory, 100GB storage) running on Proxmox 8.2.4 on an N100 (Peladn WI-6)

Continuing with the instruction of Thorsten:

pip3 install torchmetrics==0.11.4

python3.10 -m piper_train --dataset-dir ~/piper/src/python/out-train --devices 1 --batch-size 8 --validation-split 0.0 --num-test-examples 0 --max_epochs 10000 --resume_from_checkpoint ~/piper/src/python/out-train/epoch=2218-step=838782.ckpt --checkpoint-epochs 1 --precision 32 --quality high

[crispy78]

Some thoughts:
Although I really like the idea of the project, I'm really starting to dislike the project.

• There is some documentation for the piper-command, but for piper_train I couldn't find anything explaining all the arguments; Thorsten uses some that aren't even in the documentation (--quality).

• I'm getting errors I don't understand (and Google isn't very helpfull), like RuntimeError: The size of tensor a (8192) must match tensor b (7424) at non-singleton dimension 2. I see my CPU usage skyrocketing, but there aren't any checkpoints written and after several hours training was unsuccessful.

[coffeecodeconverter]

i did a guide for ubuntu 22.04 in Windows WSL
#24 (comment)