nus-cs5647-team-5/Mandarin_Tone_Evaluation

(The above metadata is required for HuggingFace to run our app)

Project Team 5 | Mandarin Tone Evaluation

Online Demo

You can try our online demo here

Before you start

create a environment using the requirements.txt file

Training

Data Preparation

You can refer to the datalist folder, which contains three subfolders: aidatatang, aishell, thchs30. Each subfolder contains corresponding label for that dataset.

You should then download the corresponding dataset from the following links:

• —Aidatatang
• —Aishell
• —Thchs30

After downloading, you can unzip the dataset and put all files into data/speech_data/ folder. For example, for Thchs30 dataset, you should put all unzipped files into data/speech_data/data_thchs30/ folder.

Model Training

run python train.py to train the model using the datasets specified in the config.json file. A typical training process will have 50 epochs, and will take about 15 hours to finish (on our Mac M1Pro chip). The model will be saved in model/ folder. We've also provided two trained model in the model/ folder to serve as the checkpoint (one trained on single dataset(thchs30), and one trained on all 3 datasets), so you can skip the training process and directly use the pretrained model to do the inference (see Inferencing section).

Inferencing

Using GUI

1. 1.Run the app.py file using command python app.py

Then you can open the given localhost link in your browser and use the GUI provided by Gradio to evaluate the tone of the given sentence.

Using CLI

1. 1.Modify the file path in predict.py to the path of the audio you want to evaluate

1. 1.Run the predict.py file using command python predict.py