Speech corpus of regional dialect variation in American English documenting changing pronunciation patterns across several generations of speakers ranging in age from 8 to 101 years old.

The Need

The speech corpus can be used for research on variation and change in spoken language, for the development of new statistical models predicting changes in pronunciation across the human lifespan, for technology applications (speech synthesis and recognition) including smartphones, robots, etc., or as a training corpus for the development of medical devices (e.g., to achieve a more natural sounding artificial voice for laryngectomy patients)

The Technology

The Corpus contains spoken utterances by 535 speakers, including the following: 264 North Carolina speakers (Southern Appalachian), 134 Wisconsin speakers (Northern variety), and 137 Ohio speakers (standard Midwestern (Midland dialect). The Corpus includes 96 children (8-14 years) and 439 adults (18-101 years), all born, raised, and spent most of their lives in the regions they represented. Seventy hours of raw speech in the Corpus includes 22k single words, 96k structured sentences, and 500 spontaneous talks.

Commercial Applications

The Corpus can be used in speech recognition technology, including speaker recognition and diarization, phonetic classification and syllabification, phonotactic modeling, pitch and prosody modeling, or speech recognition tasks that use specialized acoustic information for accent and dialect identification.

Benefits/Advantages

This cross-dialectal and the cross-generational set would contribute to developing improved algorithms for better models that capture segmental and suprasegmental variation in speech. This could also improve medical devices for patients who use artificial voice due to laryngeal cancer and surgical removal of the larynx.

The corpus provides a good training base for applying dialect-specific models due to its specific design and the common speech material collected from many speakers. The Corpus focuses on specific dialects, enabling it to systematically compare them. Also, this Corpus was created with a wider demographic, including well represented male and female speakers and ages ranging from 8-101, compared to TIMIT, which was created with mostly male speakers in their 20s and 30s. Researchers documented new developments in AE dialects, and the speakers read fluently, reducing the noise existing in other corpora.