Fig 1.
Oscillogram with Hilbert amplitude envelope for the German word Geschichte ‘history’ (top panel) and corresponding mel scaled spectogram (lower panel).
Vertical lines represent the boundaries calculated from the minima in the Hilbert amplitude envelope. For this example, 21 FBS features are extracted for each of the three chunks of speech, resulting in a total of 63 FBS features for this word.
Fig 2.
The frequency distribution of FBS features follows a power law with negative slope in the log-log plane.
Fig 3.
Examples of four realizations of German und (‘and’).
Upper left: [ʊnth], upper right: [ʊn], lowel left: [ʊnth], lower right: [n].
Fig 4.
Speaker accommodation as a function of the number of novel FBS features in held-out speech.
Each dot represents the increase in identification accuracy, comparing accuracy without and with training on the speech from the held-out speaker.
Table 1.
Coefficients, standard errors, test statistics, and p-values for the accuracy measures (upper part) and response latencies (lower table).
Fig 5.
Boxplots for the estimated by-subject coefficients for LogL1norm and LogActivation in the recognition task (upper panels) and the dictation task (lower panels).
Left: accuracy (on the logit scale); Right: latency (on the log scale). For recognition accuracy, the coefficients for LogActivation are those for the presentation over loudspeakers. For recognition latencies, the coefficients for both LogActivation and LogL1norm likewise pertain to presentation over loudspeakers.
Fig 6.
Left: Distribution of weights of afferent connections of Geschichte. Right: Identification accuracy calculated across the full data set for varying degrees of pruning.