KLEA

An open-source Khmer Word to Speech Model. Just single word not sentence!

Running With UV (Recommended)

sudo apt-get install libsndfile1 python3-dev
wget https://huggingface.co/spaces/seanghay/KLEA/resolve/main/G_60000.pth
# G_60000.pth must be in the same folder where you `uv run`
uv run --python 3.11 --with 'klea @ git+https://github.com/seanghay/KLEA' python -c 'import klea; klea.run_for_word("ទឹកធ្លាក់", "ទឹកធ្លាក់.wav")'
ffplay ទឹកធ្លាក់.wav

Installing with pip (alternative)

• Requires python >= 3.9 and python < 3.12 due to numpy and monotonic-align dependencies

sudo apt-get install libsndfile1 python3-dev
wget https://huggingface.co/spaces/seanghay/KLEA/resolve/main/G_60000.pth
pip3 install git+https://github.com/seanghay/KLEA
# G_60000.pth must be in the same folder where you run `python3`
python3 -c 'import klea; klea.run_for_word("ទឹកធ្លាក់", "ទឹកធ្លាក់.wav")'

Dataset

This model was trained on kheng.info dataset. You can find it on http://kheng.info or at https://hf.co/datasets/seanghay/khmer_kheng_info_speech

Reference

• VITS: Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech
• kheng.info is an online audio dictionary for the Khmer language with over 3000 recordings. Kheng.info is backed by multiple dictionaries and a large text corpus, and supports search in English and Khmer with search results ordered by word frequency.