Welcome to our available theses section

Are you passionate about speech recognition and artificial intelligence? Do you want to be at the forefront of cutting-edge research in Arabic language processing? Join our team at Kalam Technology, where innovation meets opportunity.

We are currently offering a unique opportunity for master's students to work on exciting thesis projects that utilize the latest advancements in transformer technologies and AI. Our projects are designed to challenge your skills, expand your knowledge, and contribute to groundbreaking developments in the field of speech recognition.

Objective:

Develop and fine-tune a transformer-based model for high-accuracy Arabic speech recognition and dialect identification

Tasks:

  • Collect and preprocess a large dataset of Arabic speech samples, including various dialects.
  • Train a transformer model (e.g., Wav2Vec 2.0 or HuBERT) for Arabic speech recognition.
  • Implement a dialect identification component and integrate it with the ASR model to improve transcription accuracy for each dialect.
  • Evaluate the combined model’s performance on different dialects and benchmark against existing systems.

Objective:

Develop a multilingual speech recognition system that excels in recognizing Arabic speech and leverages transfer learning from high-resource languages

Tasks:

  • Train a transformer-based model (like XLS-R) on a large multilingual dataset.
  • Implement strategies to improve performance on underrepresented languages, especially Arabic.
  • Use transfer learning techniques to adapt pre-trained models from high-resource languages to Arabic.
  • Assess the model’s effectiveness in multilingual and cross-lingual scenarios.

Objective:

Enhance the performance of Arabic ASR systems in low-resource and noisy environments using transformer-based models

Tasks:

  • Develop techniques to improve transformer-based ASR models’ performance on limited Arabic speech data.
  • Explore data augmentation, transfer learning, and semi-supervised learning techniques.
  • Collect and create a dataset with Arabic speech samples in various noisy conditions.
  • Train and fine-tune a noise-robust transformer-based ASR model and evaluate its performance across different noise levels.

Objective:

Develop a real-time Arabic speech-to-text translation system and a speech emotion recognition system using transformer-based models

Tasks:

  • Integrate transformer models for speech recognition (like Wav2Vec 2.0) with machine translation models (like MarianMT or mBART). .
  • Optimize the pipeline for low latency and high accuracy in translation.
  • Collect and label a dataset of Arabic speech with various emotions.
  • Train a transformer-based model to detect emotions in speech and integrate it with the ASR and translation system.
  • Evaluate the combined system’s performance in practical scenarios.

Objective:

Create a personalized transformer-based ASR system and a high-quality Arabic text-to-speech (TTS) system

Tasks:

  • Implement speaker adaptation techniques using transformer models for personalized ASR.
  • Develop a system for user-specific fine-tuning with minimal data.
  • Collect and preprocess a dataset of Arabic speech for TTS.
  • Train a transformer-based TTS model (e.g., Tacotron 2 or FastSpeech) and integrate it with the ASR system to create a complete speech interface.
  • Evaluate the personalized ASR and TTS system’s performance across different speakers.

Why Apply?

How to Apply

To apply for one of our thesis projects, please send the following to info @ kalam.se with "Thesis Application" in the subject line:

For technical questions please contact Mohammed Bakheet at cto @ kalam.se