Skip to content

Latest commit

 

History

History
105 lines (57 loc) · 3.95 KB

README.md

File metadata and controls

105 lines (57 loc) · 3.95 KB
Raw

Speaker Identification

Speaker Identification using Neural Net.

Run the file in Google Colab

Run the Beta version in Google Colab

Epochs = 10
Accuracy ~ 98%
Speakers = 3
Dataset time = 10 mins

This is a support repo for the main Project - Personalised Voice Assistant.

Make sure that you are using the same tensorflow version while trainging and running. Otherwise it will throw error ValueError: Unknown layer: Functional

ref: ValueError: Unknown layer: Functional

Data-Set

There are two options for the datasets. In either of the way you follow, download the dataset form Kaggle and keep the _background_noise_ & other folder for the Noise.

And pass the noise folder in the noise path while training and the same in the identification one too.

Kaggle

Dataset Source - Kaggle

Self Made Data-sets

We recorded the speakers voice for 10 mins using the record.py file. And the obtained 10 mins .wav is then splitted into 1 seconds of wav files using audio_clip.py.

And then append the 1 second wav files into the data-set folder with the user name.

Usage

To use the Speaker Identification from scratch, you will be needing a data set. For the testing the our model we initially trained on the Speaker recognition dataset from kaggle. Then we created our own dataset for the recognition purpose.

Dataset details:

  • time duration: 10 mins
  • splitted into 1 sec using audio_clip.py

Training

I recommend making a fresh new environment for this project.

Install the requirements by running

pip install -r requirements.txt or pip3 install -r requirements.txt

Download the Dataset or else create your own data set by speaking for 10 mins after running the file

python record.py or python3 record.py

And split the generated data.wav into 1 second files using the audio_clip.py file.
Run python audio_clip.py or python3 audio_clip.py

And train the neural net by running the the speaker-identification.ipynb file or run

python speaker_identification.py or python3 speaker_identification.py

For Regular devices

(devices with at least i3 processor or equivalent and 4 gigs of RAM)

After training, you will get the generated model.h5 locally in the root folder of the repo. And run

python predict.py or python3 predict.py

(depending upon your environment) for real time prediction. And then speak when prompted and let he

For arm/edge devices

(devices like raspberry-pi & mobile devices(Android & iOS))

For these devices you can optimize the model for them using the script tflite_convert.py, but the accuracy may reduce.

After successful training of the model you will also get model_keras_tflite.zip in your root folder of the repo.

For the conversion of the standard keras model run

python tflite_convert.py path_to_model_keras_tflite.zip or python3 tflite_convert.py path_to_model_keras_tflite.zip

And you will get a model.tflite for your arm/edge devices!

Contribution

This is a Purely open source project, and feel free to suggest changes.

To contribute, Fork it, make the changes and push a pull request and we will get back to you.

Reporting a bug

Create a issue, explaining how to reproduce the issue.

Disclaimer

We provide limited support for the tflite models.

License

Licensed under GNU General Public License v3.0