A general overview of the Voiceprint Identification project is attached Here. This paper outlines a bried introduction to the project, the teams methodology, results and discussion, and conculsions/future work. For information regarding how to use the application, view the How To Use section. A more in-depth description about how the model works is also in the How It Works setion below.
The GUI is Tkinter-based for real time classification.
Add a user profile by simply typing the user name and pressing the Add User button. The GUI will then guide you through the process of adding a user by prompting you to say a couple sentences and recording your data. If the recording time is too long or short, feel free to change the RECORD_SECONDS variable in the GUI.py file on line 125. Add a user profile by simply typing the user name and pressing the Add User button. The GUI will then guide you through the process of adding a user by prompting you to say a couple sentences and recording your data. Thats it! Now, once you train the model again to train with your data, the model can guess you. To train the model, just simply press the Train Model Button and wait until the model fully trains. Once the model has trained, the validation accuracy of the model will be updated on the UI and you wil be ready to record live!
After adding all of the neccecary user profiles and pressed the train model button, make sure to press the big red microphone button to start recording. Now, your microphone will be picking up live data, funnelling to the model, and produced a result of who is speaking. In the list of names on the left hand side of the UI, it will highlight the person who is speaking's name.
In this image, the model is 80% sure that Alex is speaking on a model that has the validation accuracy of 96.2%
Once you are all done speaking for the day, done forget to save your model before you exit to avoid retraining your data. Do this by pressing Save Model once a model has been trained. After
Data can be compiled manually or using the GUI. If data is compiled manually, the following file structure must be used: words/'word'/'name'/'file'.wav (ex words/the/john/1.wav). For an example, see the words/alexa folder. This folder contains a dataset from kaggle. The model performs well with short audio files, so break your audio data into word length files (use the function break_up_audio if you like). See the section "Add User Profile" above for how to use the GUI to add data. The GUI adds data to folder words/sentence. Note that the name of the file does not matter, and the name of the 'word' does not matter.
The word model trains a keras sequential on files in the 'words' folder. To train a model on audio data, use file structure words/'word'/'name'/'file'.wav (ex words/the/john/1.wav) and call word model as with parameter words (list of words to train on). The specific word does not matter the model trains on all the words together.
words = ["alexa","the", "be", "to", "of", "and"]
model = word_model(words)this should output
Epoch 1/50
5/5 [==============================] - 0s 63ms/step - loss: 1.7061 - accuracy: 0.3459 - val_loss: 0.7850 - val_accuracy: 0.8491
.
.
.
Epoch 50/50
5/5 [==============================] - 0s 33ms/step - loss: 0.0259 - accuracy: 0.9937 - val_loss: 0.0223 - val_accuracy: 0.9811
and make predictions by calling predict_speaker
name, confidence = predict_speaker("test/eli.wav", model)
# 'eli', .9099897this outputs the name and confidence
| Model | GUI | Misc |
|---|---|---|
| Keras | TKinter | Wave |
| Librosa | PyAudio | OS |
| Speech Recognition | PyDub | Time |
| SkLearn | Shutil | |
| Pandas | Math | |
| Tensorflow |