Date of Degree


Document Type


Degree Name



Computer Science


Andrew Rosenberg

Committee Members

Rivka Levitan

Michael Mandel

Fadi Biadsy

Subject Categories

Artificial Intelligence and Robotics


Natural Language Processing, Speech Processing


Spoken language understanding systems are error-prone for several reasons, including individual speech variability. This is manifested in many ways, among which are differences in pronunciation, lexical inventory, grammar and disfluencies. There is, however, a lot of evidence pointing to stable language usage within subgroups of a language population. We call these subgroups linguistic subcultures.

The two broad problems are defined and a survey of the work in this space is performed. The two broad problems are: linguistic subculture detection, commonly performed via Language Identification, Accent Identification or Dialect Identification approaches; and speech and language processing tasks taken which may see increases in performance by modeling for each linguistic subculture.

The data used in the experiments are drawn from four corpora: Accents of the British Isles (ABI), Intonational Variation in English (IViE), the NIST Language Recognition Evaluation Plan (LRE15) and Switchboard. The speakers in the corpora come from different parts of the United Kingdom and the United States and were provided different stimuli. From the speech samples, two features sets are used in the experiments.

A number of experiments to determine linguistic subcultures are conducted. The set of experiments cover a number of approaches including the use traditional machine learning approaches shown to be effective for similar tasks in the past, each with multiple feature sets. State-of-the-art deep learning approaches are also applied to this problem.

Two large automatic speech recognition (ASR) experiments are performed against all three corpora: one, "monolithic" experiment for all the speakers in each corpus and another for the speakers in groups according to their identified linguistic subcultures.

For the discourse markers labeled in the Switchboard corpus, there are some interesting trends when examined through the lens of the speakers in their linguistic subcultures.

Two large dialogue acts experiments are performed against the labeled portion of the Switchboard corpus: one, "monocultural" (or "monolithic") experiment for all the speakers in each corpus and another for the speakers in groups according to their identified linguistic subcultures.

We conclude by discussing applications of this work, the changing landscape of natural language processing and suggestions for future research.

This work is embargoed and will be available for download on Monday, February 01, 2021

Graduate Center users:
To read this work, log in to your GC ILL account and place a thesis request.

Non-GC Users:
See the GC’s lending policies to learn more.