Exploiting large quantities of spontaneous speech for unsupervised training of acoustic models

Purpose The study sets out to investigate inter- and intraspeaker variation in German infant-directed speech (IDS) and considers the potential impact that the factors gender, parental involvement, and speech material (read vs. spontaneous speech) may have. In addition, we analyze data from 3 time points prior to and after the birth of the child to examine potential changes in the features of IDS and, particularly also, of adult-directed speech (ADS). Here, the gender identity of a speaker is considered as an additional factor. Method IDS and ADS data from 34 participants (15 mothers, 19 fathers) is gathered by means of a reading and a picture description task. For IDS, 2 recordings were made when the baby was approximately 6 and 9 months old, respectively. For ADS, an additional recording was made before the baby was born. Phonetic analyses comprise mean fundamental frequency (f0), variation in f0, the 1st 2 formants measured in /i: ɛ a u:/, and the vowel space size. Moreover, social and behavioral data were gathered regarding parental involvement and gender identity. Results German IDS is characterized by an increase in mean f0, a larger variation in f0, vowel- and formant-specific differences, and a larger acoustic vowel space. No effect of gender or parental involvement was found. Also, the phonetic features of IDS were found in both spontaneous and read speech. Regarding ADS, changes in vowel space size in some of the fathers and in mean f0 in mothers were found. Conclusion Phonetic features of German IDS are robust with respect to the factors gender, parental involvement, speech material (read vs. spontaneous speech), and time. Some phonetic features of ADS changed within the child's first year depending on gender and parental involvement/gender identity. Thus, further research on IDS needs to address also potential changes in ADS.

Download Full-text

Perceptual Evaluation of Speech Naturalness in Speakers of Varying Gender Identities

Journal of Speech Language and Hearing Research ◽

10.1044/2020_jslhr-19-00337 ◽

2020 ◽

Vol 63 (7) ◽

pp. 2054-2069

Author(s):

Brandon Merritt ◽

Tessa Bent

Keyword(s):

Spontaneous Speech ◽

Identification Accuracy ◽

Rating Task ◽

Gender Identification ◽

Identification Task ◽

Male And Female ◽

Perceptual Evaluation ◽

Speech Training ◽

And Gender ◽

Speech Naturalness

Purpose The purpose of this study was to investigate how speech naturalness relates to masculinity–femininity and gender identification (accuracy and reaction time) for cisgender male and female speakers as well as transmasculine and transfeminine speakers. Method Stimuli included spontaneous speech samples from 20 speakers who are transgender (10 transmasculine and 10 transfeminine) and 20 speakers who are cisgender (10 male and 10 female). Fifty-two listeners completed three tasks: a two-alternative forced-choice gender identification task, a speech naturalness rating task, and a masculinity/femininity rating task. Results Transfeminine and transmasculine speakers were rated as significantly less natural sounding than cisgender speakers. Speakers rated as less natural took longer to identify and were identified less accurately in the gender identification task; furthermore, they were rated as less prototypically masculine/feminine. Conclusions Perceptual speech naturalness for both transfeminine and transmasculine speakers is strongly associated with gender cues in spontaneous speech. Training to align a speaker's voice with their gender identity may concurrently improve perceptual speech naturalness. Supplemental Material https://doi.org/10.23641/asha.12543158

Download Full-text