speech · language · machine learning

Aditya
Dalmia

MSc student at the University of Edinburgh. Building systems that listen, speak, and understand — with a background in production ASR and a growing interest in what language models can do for speech.

100k+ hours of speech data
processed at NeuralSpace
<10% WER achieved across Arabic,
Hindi, and Filipino
78 MSc sem 1 grade — Distinction
University of Edinburgh
projects
speech synthesis · 2026
Neural TTS for F1 Race Commentary
FastSpeech2-based voice cloning system trained on 2+ hours of self-recorded studio data and 20+ hours of scraped broadcast commentary. Explored curriculum learning and domain-adaptive pre-training.
audio sample coming soon
FastSpeech2 HiFi-GAN MossFormer2 pyannote
computer vision · 2026
Tissue Segmentation on PUMA Melanoma Dataset
UNet and NucleiResNet pipelines for tissue segmentation and cell classification. Autoencoder pre-training, SimCLR contrastive learning, systematic ablations. Exceeds baselines by 5–7.5%.
UNet SimCLR Lovász loss Dense CRF
asr · dissertation · ongoing
LLM-based ASR Error Correction
Investigating post-correction of ASR errors using LLMs with phoneme intermediate representations, applied to financial earnings call transcripts. In partnership with LSEG, supervised by Prof. Peter Bell.
Whisper LLM fine-tuning phonemes LSEG
advanced topics in nlp · 2026
Chain-of-Thought Reasoning with Small LLMs
LoRA fine-tuning of Qwen2.5-0.5B on GSM8K with curriculum learning, self-consistency, self-refinement, and GRPO to elicit reasoning in a sub-1B parameter model.
LoRA Qwen2.5 GRPO GSM8K
publications
ICLR 2025
INCLUDE: Evaluating Multilingual Language Understanding with Regional Knowledge
International Conference on Learning Representations, 2025
Romanou et al. · EPFL × Cohere for AI
preprint
Formula 1 Commentary Speech Synthesis in a Target Speaker’s Voice
Technical report · arXiv preprint upcoming
Aditya Dalmia
preprint
Tissue Segmentation and Cell Classification on a PUMA Melanoma Histopathology Subset
Technical report · arXiv preprint upcoming
Dalmia et al.
about

I'm a speech and language ML engineer, currently finishing an MSc at the University of Edinburgh. Before that, I spent two and a half years at NeuralSpace building production ASR systems — scaling data pipelines, training models for Arabic, Hindi, and Filipino, and shipping features used by real developers.

My dissertation looks at using LLMs to fix ASR errors, with a particular focus on named entities in financial speech — supervised jointly by Prof. Peter Bell and the London Stock Exchange Group.

Outside of work I like high-altitude Himalayan treks, Formula 1, anime, board gaming, and staring at flora and fauna around Edinburgh.

currently MSc SLP · University of Edinburgh
previously Applied Research Scientist · NeuralSpace
focus areas ASR · TTS · LLMs · Multilingual NLP
location Edinburgh, UK
languages English · Hindi · Odia
contact
✉ adi.dalmia08@gmail.com ↗ linkedin ↗ github