eman-kom/tone-sim-model

Tone Sim Model

Predicts the pinyin and tone of a student and calculates its similarity with a reference audio.

Please change the settings in config.json to accomodate to your system. Especially the filepaths. Recommended to use absolute filepaths.

Setting up the environment

conda create -n ToneSimModel
conda activate ToneSimModel
pip install -r requirements.txt

Training

To initialise the dataset: python initialise_dataset.py

To train the feature extractor: python train_classify.py

---

To train the siamese model:

With Cosine Similarity: python train_siamese.py

With Euclidean Distance: python train_siamese.py --euclid

---

To read training logs: tensorboard --logdir=runs

Testing using Examples folder

With Cosine Similarity: python -W ignore test_examples.py

With Euclidean Distance: python -W ignore test_examples.py --euclid

Inference

With Cosine Similarity: python -W ignore inference.py -r "./examples/ao1_MV1_MP3.mp3" -i "./examples/tan2_FV1_MP3.mp3"

With Euclidean Distance: python -W ignore inference.py -r "./examples/ao1_MV1_MP3.mp3" -i "./examples/tan2_FV1_MP3.mp3" --euclid