I hold both a Bachelor's and Master's degree in Electrical Engineering with a specialization in Data and Information Processing from Bar-Ilan University.
🎓 My M.Sc. thesis, conducted under the supervision of Prof. Sharon Gannot, focused on speaker separation using deep learning methods. This work led to a published paper and an open-source implementation:
📝 1."Sep-TFAnet-VAD: Joint Voice Activity Detection and Speaker Separation in Reverberant and Noisy Conditions",
EURASIP Journal on Audio, Speech, and Music Processing (2025)
The increasing complexity of real-world environments, where multiple speakers may converse simultaneously, underscores the importance of effective speech separation.
This paper presents Sep-TFAnet, a single-microphone speaker separation network based on time-frequency attention, optimized for noisy and reverberant conditions.
A variant, Sep-TFAnetVAD, jointly integrates a voice activity detector (VAD).
The model uses STFT/iSTFT in place of learned encoders and supports low-latency block processing, making it suitable for human-robot interaction.
We also introduce ARImulti-mic, a real-world dataset recorded using a humanoid robot. Project page
📝 2. "A Visual Explanation Approach for Regression Neural Networks Applied to Nearfield Acoustic Holography",
IEEE (2023)
🧠 I also co-authored a publication addressing interpretability in deep learning for regression tasks:
We propose a Grad-CAM-inspired approach to interpret neural network decisions for regression models.
The method is applied to Kirchhoff-Helmholtz-based CNN (KHCNN) for Nearfield Acoustic Holography using vibrating plates and violin top plates.
Results reveal the most informative regions used by the network for accurate predictions, and the method is validated using NCC and NMSE metrics.
Real-world audio recordings often contain transient, non-stationary noises such as keyboard clicks, door knocks, or coughs, which are particularly challenging for traditional denoising methods.
In this work, we introduce PGDI, a novel diffusion-based speech inpainting framework that reconstructs missing or corrupted speech segments by leveraging the generative power of diffusion models.
PGDI progressively denoises and reconstructs missing segments, ensuring smooth transitions and high-fidelity synthesis. The process is further guided by text-based information extracted using a language model, improving reconstruction accuracy and naturalness.
✨ Highlights:
- Progressive denoising enables smooth and natural reconstruction of both short and long gaps.
- Text guidance enhances performance for long missing segments, while the model remains effective even without text for short gaps.
- Speaker-independent — no prior knowledge of the speaker is required.
- Voice, style, and prosody preservation, ensuring realistic and natural-sounding speech.
- Environmental consistency, maintaining reverberation and acoustic characteristics.
- Strong performance on abrupt, high-energy noise scenarios such as knocks and coughs.
We evaluate PGDI with and without access to the ground-truth transcript during inference and demonstrate that text-guided PGDI excels in long-gap reconstruction, while unguided PGDI remains robust for shorter gaps.
