Fiza Husain

I’m a machine learning engineer who likes the whole loop: choosing the right question, training and adapting models to answer it, building evaluations honest enough to say when they haven’t, and then running the result in front of real users. The problems I find most interesting are in multimodal and language models, especially speech and audio, in post-training (fine-tuning, preference optimisation, adapting large models to narrow domains), and in measurement: LLM judges, metrics and experiments you can actually trust. In my experience, the number everyone optimises is often not the thing that matters. Finding that out early is most of the job.

Since June 2025 I’ve been the Founding ML Engineer at Stimuler, a conversational English-fluency platform for learners around the world, where I own the speech models end to end. My first-author INTERSPEECH 2026 paper, Beyond WER, came out of that work: it shows that entity and filler recall, not word error rate, should decide which ASR model a language-learning product uses. I also led the research behind Stimuler’s voice tutor, comparing full-duplex, cascaded and omni speech-to-speech models before shipping it on Qwen3-Omni at under 600ms to first audio. Before that I was a Research Fellow at Microsoft M365 Research, applying foundation models to incident management and intelligent monitoring for large cloud services. That work led to papers at FSE 2026 (first author), FSE 2024 and ICSE 2024, and a patent. Earlier, at IIIT Hyderabad, I worked with Prof. Praveen Paruchuri and Prof. Sujit Gujar on privacy in deep reinforcement learning, published at AAAI 2022. Alongside that, at OpenMined, I added differential privacy support to the PySyft library and fixed vulnerabilities in it.

When I’m not working, I’m usually drawing or painting, cooking, finding new places to eat, working through a maths problem for no reason at all, or chasing after a cat.

selected work

Teaching ASR to keep the words that matter

Entity- and disfluency-aware speech recognition for accented conversational English. INTERSPEECH 2026.

What actually improved a speech-to-speech tutor

Five architectures, a fine-tuning programme on Qwen3-Omni, and the finding that the largest gains came from outside the model.

Building an LLM judge you can actually trust

Validating automated evaluation against human labels, and what to do when they disagree.

Serving speech models when the metrics lie

Why GPU utilisation is the wrong signal for multi-model inference pipelines, and what to instrument instead.

latest posts

Aug 24, 2026	How do you evaluate a conversation?
Aug 24, 2026	Why word error rate is the wrong target
May 02, 2024	Probabilistic Matrix Factorization

selected publications

Beyond WER: Entity and Disfluency Recall in Accented Conversational ASR

Fiza Husain, Ankit Pandey, and Yash Singh

In Proceedings of INTERSPEECH 2026 , 2026

Abs arXiv PDF

ASR systems optimised for Word Error Rate (WER) often miss named entities and filled pauses in accented conversational English, both critical for language-learning feedback. We present a three-stage pipeline for speakers from India, Indonesia, and Latin America: (1) heuristic SQL filters curating entity-rich training data at 2.8x the entity density of random sampling, (2) regional LoRA adapters fine-tuned on Qwen2.5-Omni-3B producing both verbatim and corrected transcripts in a single forward pass, and (3) a six-category error taxonomy validated by an LLM-based judge (83.8% agreement, 210 human-labelled samples). The pipeline achieves 80-85% entity recall (up from 53-55%), 76-86% filler recall (up from <5%), and 6-10% WER across 6k test utterances, outperforming Whisper and a commercial ASR on entity recall while matching a zero-shot 30B model with 10x fewer parameters. Paired bootstrap tests confirm that curation alone accounts for 2.8-4.2 pp of entity recall gain (p<0.0001).
Attention Enhanced Entity Recommendation for Intelligent Monitoring in Cloud Systems

Fiza Husain, Anson Bastos, Anjaly Parayil, Ayush Choure, and 3 more authors

In Proceedings of the 34th ACM International Conference on the Foundations of Software Engineering (Industry Track) , 2026

Abs arXiv HTML

In this paper, we present DiRecGNN, an attention-enhanced entity recommendation framework for monitoring cloud services at Microsoft. We provide insights on the usefulness of this feature as perceived by the cloud service owners and lessons learned from deployment. Specifically, we introduce the problem of recommending the optimal subset of attributes (dimensions) that should be tracked by an automated watchdog (monitor) for cloud services. To begin, we construct the monitor heterogeneous graph at production-scale. The interaction dynamics of these entities are often characterized by limited structural and engagement information, resulting in inferior performance of state-of-the-art approaches. Moreover, traditional methods fail to capture the dependencies between entities spanning a long range due to their homophilic nature. Therefore, we propose an attention-enhanced entity ranking model inspired by transformer architectures. Our model utilizes a multi-head attention mechanism to focus on heterogeneous neighbors and their attributes, and further attends to paths sampled using random walks to capture long-range dependencies. We also employ multi-faceted loss functions to optimize for relevant recommendations while respecting the inherent sparsity of the data. Empirical evaluations demonstrate significant improvements over existing methods, with our model achieving a 43.1% increase in MRR. Furthermore, product teams who consumed these features perceive the feature as useful and rated it 4.5 out of 5.