AnoushkaAwasthi

Working out why speech models fail on accented English, and what actually fixes it.

• Work-let 26LAI11: accent-invariant representation learning for SpeechLLMs, on the robustness of speech-language models across accented English.
• Established an accent-level ASR baseline on L2-ARCTIC across Arabic, Chinese, Hindi, Korean, Spanish and Vietnamese accents.
• Evaluated Whisper Large-v3 and NVIDIA Canary-1b over 3,599 scripted utterances, with Word Error Rate as the primary metric.
• Speaker- and accent-level error analysis found that speaker variation moves WER more than accent category alone — which changes what is worth adapting.
• 24.6% of clips were misrecognised by both models, giving the next stage a concrete target rather than an average to chase.
• Now exploring parameter-efficient adaptation with LoRA on those failure cases.

Map features in a production geospatial application, in someone else’s codebase.

• Developed and extended geospatial web application features in React.js and JavaScript.
• Worked across vector and raster layers, coordinate reference systems and map projections in OpenLayers.
• Implemented and debugged map interactions and spatial visualisation in a live production codebase, which is a different job from building one from scratch.
• Worked inside an existing engineering team’s review and release workflow rather than around it.