New research on the speech features of learners of Finnish

Researchers in the DigiTala project analyzed the speech of Finnish learners in their recent study. The results on the acoustic features of non-native Finnish strengthen the automatic assessment of speech, but new questions were also raised.

Pauses and creakiness, among other features, were measured from the speech of Finnish learners who had participated in the Finnish National Certificates of Language Proficiency tests. The results show that articulation rate and duration of breaks between continuous speech runs best predict fluency ratings provided by expert raters.

Fluent speakers articulated words more rapidly and kept shorter breaks between utterances than disfluent speakers. Studies on other languages have made similar observations.

Automatic assessment measures speech features

Many automatic speech assessment systems measure features related to the speed of speech production. Similar parameters are also used in the automatic assessment tool for L2 Finnish and Swedish, which is developed in the DigiTala project.

Studying the relations between speech features and human ratings is important for the reliability of automatic assessment. Research can also discover new important features that make the automatic assessment more accurate.

Proficient speakers of Finnish creak more

The DigiTala researchers also found that some speakers, who gained higher proficiency grades, used more creaky voice than speakers with lower proficiency grades. Creaky voice is generally related to low pitch and decreasing intonation, which are typical for Finnish speech.

In Finnish, creaky voice is especially common at the end of utterances. It is possible that some advanced learners of Finnish have adopted creaky voice into their speech, consciously or unconsciously.

However, most speakers in the study did not creak at all. It is thus unclear, whether measures of creakiness can be used in improving the accuracy of an automatic assessment tool. Creaky voice is untypical for many languages and it might be difficult to adopt to one’s speech.

The results of the study will be presented in the Speech Prosody conference in May 2022.