Measuring the dynamic encoding of speaker identity and dialect in prosodic parameters

This paper describes a methodology, and the results stemming from it, for analysing the dynamic encoding of speaker identity and dialect in prosodic parameters. A method based on employing properties of the well known Dynamic Time Warping (DTW) algorithm’s path of best match allows the separation of purely dynamic from static properties of acoustic parameters and hence their evaluation as to dynamic encoding of speaker characteristics. Nineteen adult speakers of Australian English were recorded uttering a set of four sentences on five separate occasions over a period of at least one week. The prosodic parameters F0, shorttime energy, zero crossing rate and voicing were extracted for all data and analysed as to their dynamic encoding of speaker identity and dialect. Discriminate analysis (for speaker identity) and correlation analysis (for speaker dialect) analysis showed higher dynamic encoding of identity (75%) and dialect (0.58) than static encoding (55% and 0.45 respectively). Normalisation of all parameters into the range 0—1 reduced discriminate and correlation scores to 70% and 0.54 respectively. Contrasting the warp path parameters with the more conventionally employed DTW distance showed that the warp path parameters better measured speaker identity (72% versus 54%) and speaker dialect (0.56 versus 0.31) encoding. Individual analysis of the prosodic parameters shows a far higher encoding of identity and dialect in F0, though all four parameters encode dialect and identity.