AutoRPT: A Tool for Bootstrapping Prosodic Annotation

Abstract

Automated Rapid Prosody Transcription (AutoRPT) is a tool for bootstrapping manual annotation of prosodic events in either corpora or standalone audio files using the Rapid Prosody Transcription (RPT) scheme. It functions by utilizing two Long-Short Term Memory (LSTM) models, trained on measures of pitch/F0 and intensity. In addition to discrete, slightly over-generated predictions of prominence and boundary, AutoRPT produces continuous predictions between 0 and 1, similar to crowd-sourced RPT annotations averaged over listeners. Marginal predictions above a given threshold are also indicated discretely by question marks, as in the PoLaR Annotation Guidelines. Annotators achieved a statistically significant increase in annotation speed by modifying AutoRPT-generated annotations over creating annotations without assistance. In contrast with older tools such as AuToBI (Rosenberg, 2010), AutoRPT generates more theory-agnostic annotations which can support the work of non-expert annotators, and which we expect will offer greater flexibility in the prosodic annotation of other English language varieties.

Resources

Details