Fig 1.
Hierarchical rhythm structure in speech.
The speech rhythm could be hierarchically organised by either letters (a) or amplitude modulation (AM) (b) based on the speech waveform (c). Circles (a) and waveforms (b) with red, green, and blue colours represent prosody, syllable, and phoneme, respectively. Scalograms depict the AM envelopes derived by recursive application of probabilistic amplitude demodulation (d). A continuous wavelet transform was run on each AM envelope from randomly chosen 4-s speech excerpts from a neurotypical individual. The x-axis denotes time (4 s), and the y-axis denotes the modulation rate (0.1‒40 Hz). All speech signals were normalized based on z-scores (i.e., mean = 0, SD = 1).
Fig 2.
Scalograms depicting the amplitude modulation envelopes derived by recursive application of probabilistic amplitude demodulation.
A continuous wavelet transform was run on each amplitude modulation (AM) envelope from randomly chosen 4-s excerpts of AUT speech, NT speech, AUT-directed speech by NT persons, and NT-directed speech by NT persons. The x-axis denotes time (4 s), and the y-axis denotes the modulation rate (0.1‒40 Hz). The maximal amplitude is normalised to 0 dB. Note that the low frequency structure (<5 Hz) visible in NT and NT-directed speech is absent in AUT and AUT-directed speech. AUT, autism; NT, neurotypical.
Fig 3.
Overview of the interview design of the present study.
An NT questioner and either an NT or AUT respondent participated in a single interview. The NT questioner asked the AUT or NT respondent to share their recent experiences on the following 12 topics using their episodic memory: sweetness, insults, loneliness, gloom, clients, hate, transportation, kindness, agile or prompt, ginger, attractiveness, and thoughtfulness. Then, the AUT and NT respondents reported their experience on each topic. The speech by the NT questioner directed to an AUT or NT respondent was referred to as ‘AUT-directed speech’ or ‘NT-directed speech’, respectively. The speech by an AUT or NT respondent directed to the NT questioner was referred to as ‘AUT speech’ or ‘NT speech’, respectively. AUT, autism; NT, neurotypical.
Table 1.
Number of samples and mean duration of NT and AUT speech (NT and AUT respondents) and NT-directed and AUT-directed speech (NT questioners).
Fig 4.
Signal processing steps in the probabilistic amplitude demodulation (PAD) model.
Example of an amplitude modulation (AM) hierarchy derived by recursive PAD application. In the first demodulation round (left column), the data (a) are demodulated using PAD set to a fast time scale. This yields a relatively rapidly varying envelope (b) and a carrier (c). In the second demodulation round (middle column), the demodulation process is re-applied using a slower time scale than before. This yields a slower (d) and a faster (e) daughter envelope. Daughter envelopes (d) and (e) form the two levels of the resulting AM hierarchy (right column). Mathematically, these two levels (d and e) can be multiplied back by the first carrier (c, bottom left) to yield the original signal (a).
Fig 5.
Speech by individuals with autism (AUT) and neurotypical (NT) subjects.
Fast Fourier transform (a) showing a weaker prosodic rhythm in AUT speech than in NT speech, but no clear difference in syllabic or phonetic rhythms between AUT and NT. Transfer entropy analysis (b) revealed weaker top-down dynamics (e.g., prosody to syllable) of speech rhythms in the phonological hierarchy of AUT speech than of NT speech. Error bars indicate the standard deviation of the mean.
Fig 6.
Speech directed to adults with autism (AUT-directed speech) and neurotypical (NT) adults (NT-directed speech).
Fast Fourier transform (a) showing a weaker prosodic rhythm in AUT speech than in NT speech, but no clear difference in syllabic and phonetic rhythms. Transfer entropy analysis (b) showed more consistent weaker mutual dependence between different speech rhythms in AUT-directed speech than in NT-directed speech. Importantly, there is a marked similarity between AUT and AUT-directed speech and between NT and NT-directed speech (compare Fig 5). Error bars indicate the standard deviation of the mean.
Fig 7.
Correlation of frequency power of prosodic rhythm between a NT questioner and NT and AUT respondent.