Accessible Delivery of Visual-Acoustic Biofeedback for Speech Sound Disorder

Tara McAllister, Peter Traver, Amanda Eads, William Haack, Helen Carey, Yi Shan, Wendy Liang, Tae Hong ParkInterspeech 2025

TL;DR

This paper introduces staRt, an open-source iOS and web application that provides real-time visual-acoustic biofeedback for speech sound disorder (SSD) therapy. By visualizing the vocal tract's resonant frequencies (formants) as an LPC spectrum, staRt aims to make advanced speech training techniques more accessible and affordable, especially for children struggling with sounds like the American English "r".

What Problem Does This Paper Address?

Childhood speech sound disorder (SSD) can significantly impact a child's intelligibility and participation in academic and social settings. While research has shown that visual biofeedback can be highly effective for treating persistent speech difficulties, particularly for challenging sounds like the American English /r/, the high cost and specialized training required for traditional biofeedback systems have limited its widespread clinical adoption. This paper addresses the gap in accessible, evidence-based speech therapy tools by developing staRt, an application designed to democratize this technology.

Methods

The core of staRt is a real-time Linear Predictive Coding (LPC) spectrum analysis implemented in JavaScript. LPC models the vocal tract as a filter, and its coefficients are derived using the Levinson-Durbin recursion on incoming audio data. The application processes audio by disabling default browser echo cancellation and automatic gain control, optimizing parameters like buffer size and sampling rate for spectral envelope accuracy and CPU usage. A simplified peak-picking approach is used to identify formants (peaks in the LPC spectrum) that exceed a magnitude threshold, which are then visualized. For more precise formant tracking, the authors mention the potential of numerical methods like the Durand-Kerner algorithm, though they note its impact on real-time performance. The web version leverages WebRTC and the Vonage Video API for telepractice, enabling local audio analysis and real-time spectrum transmission with minimal lag.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Democratizing Advanced Therapy: The primary insight is that sophisticated biofeedback techniques, previously confined to expensive lab equipment, can be delivered effectively through readily available devices like smartphones and web browsers.
  • Visualizing the Subconscious: staRt provides a crucial visual window into the largely subconscious process of speech production, allowing individuals to compare their vocal tract resonance in real-time against a target. This is particularly helpful for those who struggle to perceive the acoustic differences between their own speech and the target sound.
  • Adaptability for Telepractice: The integration with WebRTC highlights a forward-thinking approach to modern therapy delivery, addressing the challenges of latency and audio processing inherent in remote sessions.
  • The "r" Sound as a Starting Point: Focusing on the challenging American English /r/ sound, which often persists into older childhood, demonstrates the tool's potential for addressing specific, common speech difficulties.

Key Results

The paper reports that the staRt application has been downloaded nearly 20,000 times by speech-language pathologists (SLPs) and families worldwide. While quantitative performance metrics like accuracy or latency are not explicitly detailed for the LPC spectrum analysis itself, the successful integration for telepractice with minimal lag and no loss of frequency resolution is a significant qualitative result. The application's user-facing features, including an interactive tutorial, clinician scoring, gamified reinforcement, and progress tracking, are designed to enhance user engagement and clinical utility.

Limitations and Future Work

The authors acknowledge that the current peak-picking method for formant identification is not sufficiently stable to support automated feedback features across all vocal tract sizes, especially for younger children. Future work includes expanding staRt to target other populations (e.g., English language learners) and speech sounds (e.g., sibilants like "s" and "sh"), for which a Discrete Fourier Transform (DFT) spectrum display is being prototyped. They are also exploring alternative formant estimation models, such as Kalman filtering and neural networks, to improve accuracy and enable automated feedback mechanisms, such as color-coded display changes to indicate progress towards a target. The development of a simplified display for younger children is also a stated goal.

Summary updated

BibTeX
@inproceedings{mcallister2025accessible,
  author = {Tara McAllister and Peter Traver and Amanda Eads and William Haack and Helen Carey and Yi Shan and Wendy Liang and Tae Hong Park},
  title = {{Accessible Delivery of Visual-Acoustic Biofeedback for Speech Sound Disorder}},
  booktitle = {Interspeech},
  year = {2025},
}

Loading related content...

Figure