Research intern at the EPFL NLP Lab in Lausanne, working with Antoine Bosselut on native audio understanding for Apertus 2, Switzerland's fully open multilingual LLM.
I work on multilingual and low-resource NLP, mostly around speech. My research measures how far synthetic (TTS) speech can substitute for real data in low-resource languages, and builds the corpora and benchmarks that Persian NLP has been missing — including ParsVoice, the largest open Persian speech corpus (3,500+ hours), created for my M.Sc. thesis at the University of Tehran (GPA 4.0/4.0) under Azadeh Shakery and Heshaam Faili.
Hear my work بشنوید
Zero-shot voice cloning for Persian: give the model a few seconds of a voice it has never heard, plus a sentence — it speaks the sentence in that voice. This line was never recorded by a human — it's my thesis model, cloning a voice from the ParsVoice demo:
صدای خنده بچهها در تمام کوچه پیچیده بود ▶ press to listen
"The sound of children's laughter echoed throughout the alley."
Full demo · dataset (3,500+ hours) · paper
That covers the generation side. On the understanding side, I built PARSA-Bench — 16 audio-language tasks for Persian, 10 of them the first of their kind in any language — and at EPFL I'm now bringing audio and text together inside one open multilingual LLM.
Experience تجربه
Audio understanding for Apertus 2; exchange-rate and training-curriculum studies quantifying the value of synthetic speech across five low-resource languages.
Led research on Persian speech synthesis and medical LLMs; chief TA for ML Methods in NLP and Digital Speech Processing; TA for five more courses serving 200+ students.
ML and deep learning courses: built an LLM-agent assignment and organized an ML contest with 200+ participants.
Multi-platform RAG customer-service chatbot (Telegram, WhatsApp, web); 20,000+ active users on the University of Tehran website.
Mentored international students in deep learning with a focus on NLP.
Taught image processing and ML; developed the SWIMBot platform for Diginext; research on facial emotion recognition.
Selected Publications مقالهها
All publications →Selected Projects پروژهها
All projects →Interactive demo of zero-shot voice cloning for Persian, built on the ParsVoice corpus.
Low-cost robot with real-time facial emotion recognition, gesture responses, and speech commands.
Multi-platform chatbot with a custom RAG pipeline; 20,000+ active users at the University of Tehran.
Education تحصیلات
GPA 4.0/4.0 · Thesis: Zero-Shot Text-to-Speech for the Persian Language, advised by Azadeh Shakery and Heshaam Faili.
Thesis: Facial Expression Recognition with Robotic Implementation, advised by M. B. Menhaj and H. Taheri.