Language science & language technology
From fieldwork to production systems.
I am engaged in two kinds of research: language science and language technology. I love to help extend our collective understanding of both how human interaction works and what computational systems can usefully do with speech and language.
Language science
I've mostly been interested in language as a complex system shaped by the social and cognitive constraints of human interaction and learning. We can model aspects of speech and language as mathematical or computational systems, while holding lightly to these abstractions and tightly to the data details, knowing the abstractions are incomplete. The full publication list is on Google Scholar.
Stress and tone
-
Lexical and post-lexical tone in Choguita Rarámuri
-
Tone and stress in Choguita Rarámuri word prosody
Caballero & Carroll. International Journal of American Linguistics, 2015.
-
Ixpantepec Nieves Mixtec word prosody
Some historical context for these papers is in the language documentation research archive.
Phonological typology
-
Contrasts and contexts in palatalization
Datasets
-
Accessing, managing and mobilizing an ELAN-based language documentation corpus: the Kwaras and Namuti tools
Caballero, Carroll & Mach. Language Documentation and Conservation 13: 63–82.
-
Materials of the Choguita Rarámuri Language Project
Chaparro Gardea et al. SCL 2019-01, Survey of California and Other Indian Languages.
Language technology
Since my first computational linguistics class more than 25 years ago, computational systems have become much more useful for processing speech and language, and yet the path from an experimental prototype to a reliable production system still depends on a combination of user-centered design, pragmatic engineering, and rigorous evaluation.
Conversational interfaces
-
Developing production-level conversational interfaces with shallow semantic parsing
-
Entity resolution for noisy ASR transcripts
Raghuvanshi, Ramakrishnan, Embar, Carroll & Raghunathan. EMNLP 2019.
-
Automatic speech recognition correction
Raghunathan, Raghuvanshi, Ramakrishnan, Carroll & Embar. US Patent 11,482,213 (2022); US Patent 12,300,225 (2025, continuation).
Multilingual ASR
-
XLSR/XLS-R fine-tuning for low-resource languages
Speech models fine-tuned for Kinyarwanda, Luganda, Uzbek, Uyghur, and Kyrgyz. Hugging Face community events.
Input methods
-
Koreguaje keyboard
Topic segmentation
-
Evaluating hierarchical discourse segmentation
Some historical context of this work is in the segmentation research archive.
Question answering
-
Cohere Labs Community at FoodBench-QA 2026: The cake makes the ingredients
Ranjan, Santhosh & Carroll. CL4Health Workshop at LREC 2026.
-
Kedara Guide
RAG-based chat assistant for finding eldercare resources.
These are the same problems I take on in consulting: evaluation design, data quality, and failure analysis.
Working on a hard language problem?
I am especially interested in collaborations about multilingual speech technology for healthcare applications, and collaborations with indigenous communities about custom language technology.