Maria Teleki

📄 buff.ly/S0DSZzt
⚽️ Xiangjue Dong (1st author), Cong Wang, Millenium Bismay, and James Caverlee
#NLP #NLPResearch #LLMs #GenAI #AI

🌟 To me, this work is super exciting because we take a totally different perspective: we show that ⬆️ diverse perspectives, ⬆️ system performance, so ⬆️ $$$ for a company! With this work, we argue that <<< 🚨 diverse perspectives are absolutely necessary >>> from an economic standpoint.

You always hear about the "bias-accuracy tradeoff," meaning that ⬇️ model bias, ⬇️ system performance, so ⬇️ $$$ for a company. So much of the conversation around bias and diversity has focused on how to incentivize companies to debias their models (e.g., through new legislation).

In our new work, 🎶 CHOIR: Collaborative Harmonization fOr Inference Robustness, we show that different LLM personas often get different benchmark questions right! CHOIR leverages this diversity to boost performance across benchmarks. 📊

LLMs can write — but can they tell stories?

Our survey shows they struggle with:
⚠️ Long-term coherence
⚠️ Controllability

📚 Paper: mariateleki.github.io/pdf/A_Survey...

#StoryGeneration #GenerativeAI #NLP

Choosing an ASR system isn’t one-size-fits-all — it depends on the disfluencies in your domain.

📄 www.isca-archive.org/interspeech_...

Disfluencies aren’t just noise — they’re part of how we speak.

In our #INTERSPEECH2024 paper, we looked at how Google ASR vs WhisperX handle messy, real-world podcasts (82k+ episodes!):

🎙️ WhisperX → better with “uh/um”
📝 Google ASR → better with self-corrections

Highlight of my PhD → mentoring students. It's literally just the most fun to brainstorm with them each week and watch them learn and grow 🌱

#AcademicMentoring #PhDLife

We can’t fix what we don’t measure.

That’s why I build evaluation frameworks for speech & conversational AI — so we can stress-test systems against real-world variability.

#AIResearch #Evaluation #SpeechProcessing

I’m working on methods & evaluation frameworks for conversational AI that are:
✅ Robust to disfluencies
✅ Reliable in noisy, real-world conditions
✅ Generalizable across contexts

If conversational AI is going to truly work for everyone, it must be built for human speech as it is.

Why “speech-first” AI?
Because speech ≠ text.

🎙️ People pause, restart, self-correct
🌎 Background noise & accents vary
💬 Context shifts across domains

What happens when you say:
“I want a horror -- comedy -- movie”? 🎥

That slip-of-the-tongue can confuse recommender systems.
Our INTERSPEECH 2025 paper shows some LLMs handle it better than others.

📄 mariateleki.github.io/pdf/HorrorCo...

#INTERSPEECH2025 #ConversationalAI #RecSys

These insights still apply to anyone working on conversational AI, spoken summarization, or voice-driven interfaces today.

📄 Read more: www.isca-archive.org/interspeech_...

#SpeechProcessing #ConversationalAI #VoiceAI #Disfluency #SpokenLanguage
#INTERSPEECH

We compared two systems on 82,000+ podcast episodes. We found:
👉 WhisperX better captures interjections like “uh” and “um”
👉 Google ASR better captures edited nodes (e.g., “let’s go to Target--Walmart”)

🌟 The type of disfluency matters when choosing an ASR system

Last year at INTERSPEECH 2024, we explored a question that remains relevant: how do ASR systems handle disfluencies in real-world speech?

Stories shape how we think and connect. 📖
But can AI tell a good story?

Our Survey on LLMs for Story Generation (EMNLP Findings 2025) explores:
✨ Coherence
🎛️ Controllability
🎨 Creativity
⚖️ Authenticity

📄 mariateleki.github.io/pdf/A_Survey...

#StoryGeneration #GenerativeAI

Speech isn’t perfect.
We restart, repeat, and slip.

For AI, those little disfluencies can cause big problems.
That’s why my research builds methods to make spoken language systems more robust.

#SpeechProcessing #ConversationalAI #NLP #AI

📄 Paper: mariateleki.github.io/pdf/A_Survey... | 💻 GitHub: github.com/mariateleki/...

#StoryGeneration #GenerativeAI #NLProc

Back in August, we shared our Survey on LLMs for Story Generation (EMNLP Findings 2025).

📚 Covers: controllability, coherence, and creativity
🧩 Discusses: evaluation challenges
🌍 Highlights: hybrid symbolic–neural approaches
💻 Includes: an open resource list (PRs welcome!)

#arxiv #speechAI #LLM

🧭 9 actionable recommendations for deployment -- some are surprising! 😱
📄 Paper: arxiv.org/pdf/2509.20321
💻 Code: github.com/mariateleki/...

🔬 DRES provides the first controlled benchmark for evaluating LLMs on disfluency removal.
✅ Controlled evaluation on gold transcripts (no ASR noise) sets an upper bound
📊 Systematic comparison across open & proprietary LLMs
🧪 First taxonomy of LLM error modes

🚀 New on arXiv: We introduce DRES, the disfluency removal evaluation suite!

Z-Scores: A Metric for Linguistically Assessing Disfluency Removal Evaluating disfluency removal in speech requires more than aggregate token-level scores. Traditional word-based metrics such as precision, recall, and F1 (E-Scores) capture overall performance but…

🔎 Z-Scores reveal model weaknesses by disfluency type — EDITED, INTJ, and PRN — providing diagnostic insights that guide targeted improvements.

📄 Paper: arxiv.org/abs/2509.20319
💻 Code: github.com/mariateleki/...

🌟 New on arXiv — we introduce Z-Scores: A Metric for Linguistically Assessing Disfluency Removal 📊🧠

🤔 Traditional F1 scores hide why disfluency removal models succeed or fail.

🌱 The takeaway: model selection is critical for real-world conversational AI.

📄 Full paper & code: mariateleki.github.io/pdf/HorrorCo...

#SpeechProcessing #ConversationalAI #VoiceAI #Disfluency #SpokenLanguage
#INTERSPEECH #ICASSP

We tested multiple LLMs as CRS backbones and found:

🌟 Llama and Mixtral became more resilient with these errors
🌟 Gemini, GPT-4o, GPT-4o-mini performance dropped

Disfluencies aren’t just noise — they can even help certain models by introducing genre diversity.

Speech is messy — and so are recommender systems when they face speech errors.

In our INTERSPEECH 2025 paper, we introduced Syn-WSSE, a psycholinguistically grounded framework for simulating whole-word substitution errors in conversational recommenders (e.g., “I want a horror—comedy movie”).

📄 Paper/code: www.isca-archive.org/interspeech_...

#SpeechProcessing #ConversationalAI #VoiceAI #Disfluency #SpokenLanguage

Messy, real-world speech ≠ clean transcripts.

In our #INTERSPEECH2024 paper, we compared Google ASR vs WhisperX on 82k+ podcasts 🎙️

🌱 WhisperX → better with accurately transcribing “uh/um”
🌱 Google ASR → better with accurately transcribing edited nodes
🌱 Which to use? Depends on your data.