Generates samples using Piper for training a wake word system like openWakeWord.
Available models:
Create a virtual environment and install the requirements:
git clone https://github.com/rhasspy/piper-sample-generator.git
cd piper-sample-generator/
python3 -m venv .venv
source .venv/bin/activate
python3 -m pip install --upgrade pip
python3 -m pip install -r requirements.txt
Download the LibriTTS-R generator (exported from checkpoint):
wget -O models/en-us-libritts-high.pt 'https://github.com/rhasspy/piper-sample-generator/releases/download/v2.0.0/en_US-libritts_r-medium.pt'
See links above for models for other languages.
Generate a small set of samples with the CLI:
python3 generate_samples.py 'okay, piper.' --max-samples 10 --output-dir okay_piper/
Check the okay_piper/
directory for 10 WAV files (named 0.wav
to 9.wav
).
Generation can be much faster and more efficient if you have a GPU available and PyTorch is configured to use it. In this case, increase the batch size:
python3 generate_samples.py 'okay, piper.' --max-samples 100 --batch-size 10 --output-dir okay_piper/
On an NVidia 2080 Ti with 11GB, a batch size of 100 was possible (generating approximately 100 samples per second).
Setting --max-speakers
to a value less than 904 (the number of speakers LibriTTS) is recommended. Because very few samples of later speakers were in the original dataset, using them can cause audio artifacts.
See --help
for more options, including adjust the --length-scales
(speaking speeds) and --slerp-weights
(speaker blending) which are cycled per batch.
Alternatively, you can import the generate function into another Python script:
from generate_samples import generate_samples # make sure to add this to your Python path as needed
generate_samples(text = ["okay, piper"], max_samples = 100, output_dir = output_dir, batch_size=10)
There are some additional arguments available when importing the function directly, see the docstring of generate_sample
for more information.
Once you have samples generating, you can augment them using audiomentation:
python3 augment.py --sample-rate 16000 okay_piper/ okay_piper_augmented/
This will do several things to each sample:
impulses/
Models that were trained on audio books tend to perform poorly when speaking short phrases or single words. The French, German, and Dutch models trained from the MLS have this problem.
The problem can be mitigated by repeating the phrase over and over, and then clipping out a single sample. To do this automatically, follow these steps:
<phrase>,
)--noise-scales 0.333
and --noise-scale-ws 0.333
--min-phoneme-count 300
(the value 300 was determined empirically and may be less for some models)For example:
python3 generate_samples.py \
'framboise,' \
--model models/fr_FR-mls-medium.pt \
--noise-scales 0.333 \
--noise-scale-ws 0.333 \
--min-phoneme-count 300
--max-samples 1 \
--output-dir .