Can We Predict Upcoming Stuttering Moments?

Imagine that there’s a word you want to say but it just isn’t coming out – this frustrating feeling is similar to what many people who stutter (PWS) experience in daily life.

🗣What is stuttering?

Stuttering is a speech fluency disorder characterised by involuntary speech disruptions, such as:

• Part-word repetitions – “Th-th-this is”
• Blocks – “a (um)”
• Prolongations of sounds or syllables – “blog ppppost.”

Stuttering commonly appears as a developmental disorder, with an average age of onset between 2 to 4 years old. It can also persist into adulthood, in which cases, known as persistent developmental stuttering (Bloodstein & Bernstein Ratner, 2008). This affects approximately 1% of the adult population (Yairi & Ambrose, 2013).

🦻Delayed Auditory Feedback as a treatment approach for stuttering

While no complete cure for stuttering has been identified yet, many modes of treatment have been found to effectively reduce stuttering and improve fluency.

One of the most effective and commonly used fluency enhancement methods is delayed auditory feedback (DAF). DAF is defined as when electronic changes cause speakers to experience a delay in hearing their own voice of what they just said, usually by 100-200ms (Goldiamond, 1965). This is used in stuttering-reduction devices like SpeechEasy.

However, maladaptive side effects after constant use of DAF have also been reported, including:

• prolongation of vowels (Howell et al., 1988).
• reduced speech naturalness (Stuart & Kalinowski, 2004).

One solution to potentially tackle these limitations and maximise the benefits of DAF, would be to restrict the delayed feedback to only moments of stuttering, instead of one’s entire speech output.

To do so would require accurate prediction of upcoming stutters.

🧠Can we ‘predict’ stuttering moments?

While the idea of trying to predict something that hasn’t happened yet sounds extremely difficult, neurological evidence from abnormalities in speech motor preparation in PWS signal the possibility to predict upcoming stutters. Studies have found that PWS show larger motor preparation effort as a compensation to produce fluent speech (Vanhoutte et al., 2016).

Hence, we believe that there are certain underlying biomarkers to help predict upcoming stutters. However, it is almost impossible to promote the use of these neurophysiological (e.g. EEG) biomarkers to everyday usage. It is definitely too expensive, not widely accessible, and has certain ethical and safety issues.

💬Current study: finding an acoustic signal to predict stutters

Hence, our current question is: Are there any voice features before stuttering caused by abnormal motor preparation, that can be perceived by the human ear? If so, based on how the human ear perceives these features, computational models can learn to predict stuttering patterns, and eventually automate this process. Howell and Wingfield (1990) discovered that naive listeners were able to reliably differentiate which sections of speech had been near a stutter, and even whether the stutter was of repetition or prolongation type.

Our current study is aimed to investigate whether listeners can reliably predict an upcoming stutter from the preceding fluent parts of speech by ear. We hypothesised that, if sufficient acoustic signals of abnormal motor preparation are presented, listeners should be able to predict upcoming stutters, and give correct “stuttered/fluent speech is coming up” responses according to the seemingly fluent speech before the stutter.

🔎Methods

Participants: 40 neuro-typical participants aged 18-30, with normal hearing, and native/fluent in English.

Design

• Experimental condition: fluent speech segment which was originally followed by a stutter.
• Control condition: fluent speech segment originally followed by fluent speech.
• Dependent Variable: Judgement of stutter likelihood on scale of 1-7.

Materials: 88 audio clips were extracted from archived spontaneous speech data [UCLASS] of stuttering speakers, using SFS and Audacity. Audios in the experimental condition were cropped right before stutter onset. All speech segments were grouped by similar syntactic and phonological difficulty.

Audio examples from our study: Can you spot any differences?

Procedure:

📝Results

A two sample t-test was conducted to compare the stutter prediction scores in the two conditions. Within the scale of 1 to 7, with 7 being the highest likelihood of a stutter to occur, the participants on average gave significantly higher ratings of predicted stutter likelihood in the experimental (actually stuttered) condition (Mean rating = 3.37) compared to the control condition (Mean rating = 3.04).

In addition, text responses to the question ‘Were there any cues that helped you decide whether the following speech is stuttered?’ were analysed. Speech Rate was the most identified feature (mentioned 12 times), followed by Pauses and Hesitations (9 times) and Clarity (3 times). Example quotation: “Whether the speakers pause in the middle or whether they speak slowly.”

✅Implications and conclusion

1. The answer to our question is YES! Naive listeners were able to detect differences in the fluent speech before stuttered and fluent speech. These decisions were frequently based on perceived differences in speech rate, number/ length of pauses, and speech clarity etc.
2. However, even for the experimental (stuttered) condition, stutter likelihood ratings were lower than 4 – the midpoint of the rating scale. This is not surprising, because many of the mentioned speech features might be too subtle for the human ear to reliably detect, making the task considerably difficult. Computational models however, might be able to more accurately capture these acoustic features.

In conclusion, this study provides perceptual evidence for biomarkers that enable upcoming stutter prediction. Future research could control and examine the effects of different acoustic features (e.g. speech rate).

The public, especially friends and family of people who stutter can learn more about stuttering from this study. More importantly, it could help develop computational methods towards automated DAF devices, and other generalizable clinical applications, which greatly help people who stutter enhance speech fluency.

References

Bloodstein, O., Ratner, N. B., & Brundage, S. B. (2021). A Handbook on Stuttering, Seventh Edition. Plural Publishing.

Goldiamond, I. (1965). Stuttering and fluency as manipulatable operant response classes. Research in Behavior Modification. https://cir.nii.ac.jp/crid/1573105975417659264

Howell, P., & Wingfield, T. (1990). Perceptual and acoustic evidence for reduced fluency in the vicinity of stuttering episodes. Language and Speech, 33 ( Pt 1), 31–46. https://doi.org/10.1177/002383099003300103

Howell, P., Wingfield, T., & Johnson, M. (1988). Characteristics of the speech of stutterers during normal and altered auditory feedback. Institute of Acoustics.

Stuart, A., & Kalinowski, J. (2004). The perception of speech naturalness of post-therapeutic and altered auditory feedback speech of adults with mild and severe stuttering. Folia Phoniatrica et Logopaedica: Official Organ of the International Association of Logopedics and Phoniatrics (IALP), 56(6), 347–357. https://doi.org/10.1159/000081082

Vanhoutte, S., Cosyns, M., van Mierlo, P., Batens, K., Corthals, P., De Letter, M., Van Borsel, J., & Santens, P. (2016). When will a stuttering moment occur? The determining role of speech motor preparation. Neuropsychologia, 86, 93–102. https://doi.org/10.1016/j.neuropsychologia.2016.04.018

Yairi, E., & Ambrose, N. (2013). Epidemiology of stuttering: 21st century advances. Journal of Fluency Disorders, 38(2), 66–87. https://doi.org/10.1016/j.jfludis.2012.11.002