Language Identification using CNN PyTorch

Language and Libraries

Problem statement

The goal of this project is to develop a Language Identification system that can accurately identify the spoken language from an audio file. The system should be able to process audio files in various formats, such as .mp3, and predict the language with high accuracy.

Solution Proposed

To address this problem, we proposed a solution that utilizes the Pytorch framework for machine learning and the Torchaudio library for audio processing. The audio is first converted to a Mel Spectrogram, which is then fed into an image classification model. We created a custom Language Identification network using Pytorch, which is trained on a dataset of audio samples in various languages.

The system was tested on a set of audio samples and was able to achieve high accuracy in identifying the spoken language. We also developed an API that takes in an audio file in the form of an .mp3 file and returns the predicted language. Finally, the application was containerized using Docker and deployed on the AWS cloud for easy access and scalability.

Dataset Used

This is a dataset of audio samples of 4 different Indian languages. Each audio sample is of 5 seconds duration. This dataset was created using regional videos available on YouTube.

This is constrained to Indian Languages only but could be extended.

Languages present in the dataset - Hindi, Kannada, Tamil, Telugu.

Points to Improve:

• Incorporation of more diverse dataset of audio samples from different languages and accents to improve the accuracy of the model.

• Experimentation with different audio processing techniques to see if it improves the performance of the model.

• Integration of the system with a user-friendly interface for easier access and usage for non-technical users.

• Further optimization of the model to reduce the latency of the API and improve the overall efficiency of the system.

• Incorporation of additional features such as speaker identification and text transcription to make the system more versatile.

• Evaluate the performance of the model with different languages and accents, and if needed retrain the model with more diverse dataset.

How to run?

Step 1: Clone the repository

git clone "https://github.com/aravind-selvam/Language_Identification-using-CNN-and-Audio-Processing.git" repository

Step 2- Create a conda environment after opening the repository

conda create -p env python=3.10 -y

conda activate env/

Step 3 - Install the requirements

pip install -r requirements.txt

Step 4 - Export the environment variable

export AWS_ACCESS_KEY_ID=<AWS_ACCESS_KEY_ID>

export AWS_SECRET_ACCESS_KEY=<AWS_SECRET_ACCESS_KEY>

export AWS_DEFAULT_REGION=<AWS_DEFAULT_REGION>

Before running server application make sure your s3 bucket is available and empty

Step 5 - Run the application server

python app.py

Step 6. Train application

http://localhost:8080/train

Step 7. Prediction application

http://localhost:8080

Run locally

1. Check if the Dockerfile is available in the project directory

2. Build the Docker image

docker build -t langapp .

3. Run the Docker image

docker run -d -p 8080:8080 <IMAGEID>

👨‍💻 Tech Stack Used

1. Python
2. Flask
3. Pytorch
4. Docker
5. CNN

🌐 Infrastructure Required.

1. AWS S3
2. GAR (Google Artifact repository)
3. GCE (Google Compute Engine)
4. GitHub Actions

src is the main package folder which contains

Artifact : Stores all artifacts created from running the application

Components : Contains all components of Machine Learning Project

• DataIngestion
• DataTransformation
• ModelTrainer
• ModelEvaluation
• ModelPusher

Custom Logger and Exceptions are used in the project for better debugging purposes.

Conclusion

This Language Identification system can be used in a variety of applications such as:

• Speech-to-text transcription: The system can be used to automatically identify the language of the audio input and transcribe it to text in the corresponding language, improving the accuracy of the transcription.
• Contact centers: The system can be integrated into contact center systems to automatically identify the language spoken by the customer and route the call to an appropriate agent.
• Language identification in the field of speech recognition: The system can be used to identify the spoken language of the speaker, thus allowing speech recognition systems to adapt to the speaker's accent and language.
• Audio/Video Translation: The system can be used to identify the language of the audio/video, then use the language to translate the audio/video in real-time.

=====================================================================