New release of Mixer-6: Improved validity for phonetic study of speaker variation and identification

New Release of Mixer-6: Improved Validity for Phonetic Study of Speaker Variation and Identification Eleanor Chodroff, Matthew Maciejewski, Jan Trmal, Sanjeev Khudanpur, John Godfrey Johns Hopkins University 3400 N. Charles St., Baltimore, MD 21218 echodro1@jhu.edu, mmaciej2@jhu.edu, jtrmal@gmail.com, khudanpur@jhu.edu, godfrey.jack@gmail.com Abstract The Mixer series of speech corpora were collected over several years, principally to support annual NIST evaluations of speaker recognition (SR) technologies. These evaluations focused on conversational speech over a variety of channels and recording conditions. One of the series, Mixer-6, added a new condition, read speech, to support basic scientific research on speaker characteristics, as well as technology evaluation. With read speech it is possible to make relatively precise measurements of phonetic events and features, which can be correlated with the performance of speaker recognition algorithms, or directly used in phonetic analysis of speaker variability. The read speech, as originally recorded, was adequate for large-scale evaluations (e.g., fixed-text speaker ID algorithms) but only marginally suitable for acoustic-phonetic studies. Numerous errors due largely to speaker behavior remained in the corpus, with no record of their locations or rate of occurrence. We undertook the effort to correct this situation with automatic methods supplemented by human listening and annotation. The present paper describes the tools and methods, resulting corrections, and some examples of the kinds of research studies enabled by these enhancements. Keywords: corpus, speech, transcription 1. Introduction Speech corpora for research on speaker recognition (SR) are particularly challenging to design and collect. Best practices in this field call for accurate demographics, multiple sessions per speaker at certain intervals, control of speech style, signal quality, channels, transducers, and many other factors, all with accurate documentation. The NIST Speaker Recognition evaluations (SRE) have provided de facto standards for such corpora over the last two decades, and the Mixer series of corpora were collected to meet SRE requirements over a period of several years (Cieri et al., 2004). They exercised a variety of conditions relevant to their intended applications: multiple languages, channels, environments, noises, speaking styles and many others. Mixer-6 followed the same protocols as the other Mixers in the collection of phone calls, but added two additional recording conditions of an interview and read speech portion (for a detailed description of the design and methods, see Brandschain et al., 2013). For forensic, phonetic, and engineering research, this corpus can be very valuable, especially with regards to speaker identification and variability. In comparison to other spoken corpora, some of the distinguishing features of the Mixer-6 read speech include the sheer number of unique speakers and the amount of data collected per speaker, multiple recording sessions for most speakers, and the controlled sentence content and order. This design was primarily intended to advance research in forensic applications, i.e., SR decisions made by human experts for evidentiary use, with or without the aid of computer algorithms. The goal of this paper is to present an improved, audited version of the Mixer-6 read speech and to highlight its potential not only for forensic, but also for other …

Content truncated to 500 words