Flowen: Real-Time Acoustic Biofeedback for Disfluency — Technical & Clinical Overview

1. Abstract

Flowen is a cloud-based speech-technology platform that delivers real-time acoustic biofeedback to individuals who stammer, their speech and language therapists (SLTs), and funding bodies including NHS Integrated Care Boards and the Access to Work (AtW) scheme. The platform combines a proprietary disfluency-detecting automatic speech recognition (ASR) pipeline with a dual-waveform visualisation method to provide moment-by-moment feedback on fluency events — blocks, prolongations, and repetitions — during self-directed or clinician-guided practice sessions.

This paper describes the scientific rationale for acoustic biofeedback in stammering therapy, the architecture of the Flowen ASR system, the platform's data-privacy and clinical-safety governance, and the intellectual-property landscape. It is intended for clinicians, commissioners, investors, and technical reviewers evaluating the platform.

2. Introduction

Stammering (also called stuttering) is a neurologically based speech disorder affecting approximately 1% of adults worldwide. Despite decades of established therapeutic approaches, access to consistent, measurable practice is a persistent barrier: waiting lists for NHS SLT services frequently exceed 12 months, self-directed practice lacks objective feedback, and progress is difficult to quantify for commissioning purposes.

Flowen was founded in July 2026 by a person who stammers — someone who has lived with the condition for 29 of his 31 years. That founding context is not incidental; it shapes every product decision. The platform was built first as the tool the founder needed and could not find: objective, on-demand feedback available outside the clinic, without a waiting list, without expensive hardware, and without the self-consciousness of practising in front of another person. Flowen deploys consumer-grade device hardware — a smartphone or laptop microphone — and specialist deep-learning models to deliver the kind of acoustic feedback previously available only in specialist clinical settings or expensive hardware devices (e.g. SpeechEasy® delayed auditory feedback devices). By making biofeedback accessible at home, at work, and on-demand, Flowen aims to compress the gap between clinic-guided therapy and independent practice.

3. Clinical Rationale

3.1 Prevalence & Burden of Stammering

Stammering affects an estimated 70 million people globally (Bloodstein & Ratner, 2008) and approximately 700,000 adults in the United Kingdom (STAMMA, 2023). The condition carries significant psychosocial burden: adults who stammer report higher rates of anxiety, social avoidance, unemployment, and reduced quality of life compared with fluent speakers (Craig et al., 2009; Briley & Ellis, 2018).

NHS England's 2022 Service Specification for Stammering recognises the need for blended delivery models, including digital support, to improve access — particularly for adults in underserved regions where specialist SLT provision is thin.

3.2 Evidence Base for Acoustic Biofeedback

Acoustic biofeedback — the real-time presentation of speech signal characteristics to the speaker — has been investigated as a stammering intervention since the 1960s. The strongest evidence base centres on three modalities:

• Delayed Auditory Feedback (DAF):Presenting the speaker's own voice with a delay of 50–200 ms induces slowed, more fluent speech in many speakers (Lincoln et al., 2006; van Borsel et al., 2003). Hardware DAF devices (e.g. SpeechEasy®) achieve clinically meaningful fluency reductions in controlled settings.
• Frequency-Altered Feedback (FAF):Shifting the pitch of the speaker's feedback signal by ±0.5 semitones reduces stuttering frequency in 50–70% of participants in laboratory conditions (Stuart et al., 2004).
• Visual Biofeedback (VBF): Real-time acoustic visualisations (pitch tracks, spectrograms, and waveform displays) facilitate technique acquisition in fluency-shaping programmes, particularly Easy Onset (smooth voicing onset) and stretched speech (Onslow et al., 2017; Packman et al., 2014). VBF enables speakers to observe their own vocal patterns and self-correct in real time.

Flowen focuses primarily on Visual Biofeedback, augmented with automated disfluency event markers. This approach is supported by a growing body of evidence suggesting that self-monitoring via visual display enhances generalisation of fluency techniques beyond the clinic (Euler et al., 2014; Cream et al., 2019).

4. Technology Architecture

4.1 ASR Pipeline Overview

Flowen's ASR pipeline processes microphone audio in near-real-time (target latency <300 ms end-to-end). Audio is captured at 16 kHz, 16-bit PCM and processed in overlapping frames of 25 ms with a 10 ms hop. The pipeline consists of three sequential stages:

Audio is processed entirely on-device in the browser using WebAssembly where possible, falling back to server-side inference for users on lower-powered devices. No raw audio is transmitted to Flowen servers unless the user explicitly consents to session recording for their own progress review.

4.2 Disfluency Detection Method

The core detection model is a transformer-based encoder fine-tuned on Flowen's proprietary disfluent speech corpus (described in §5). The model outputs three event types at the sub-word level:

• Blocks:Silent or near-silent fixations at word onset — detected by anomalous silence duration preceding a phoneme onset. Threshold: >200 ms pre-vocalic silence during voiced speech.
• Prolongations: Abnormal extension of a vowel or fricative beyond expected phoneme duration. Detected via per-phoneme duration z-score against speaker-adapted norms.
• Repetitions: Whole-word or part-word repetitions, detected via n-gram matching on the token sequence with a minimum repetition window of 3 tokens.

At the time of writing, the model achieves the following performance on the Flowen internal test set (500 sessions, held-out speakers):

4.3 Biofeedback Rendering — Dual Waveform

Flowen renders two simultaneous waveforms on a shared timeline. The amplitude waveform (top channel) displays the raw speech signal, allowing users to observe breath support, loudness variation, and silent periods. The formant waveform(bottom channel) tracks first-formant (F1) energy as a proxy for degree-of-voicing and vowel openness, providing visual cues for Easy Onset technique acquisition.

Disfluency events are overlaid as coloured markers: amber for blocks, violet for prolongations, and rose for repetitions. A real-time fluency score (0–100) is derived from the weighted sum of event counts per minute and displayed as a gauge for immediate session feedback.

The dual-waveform visualisation approach is a proprietary method currently under assessment for patentability (see §8).

5. Data & Privacy

Flowen is built to the UK GDPR standard. Lawful basis for processing session data is Article 6(1)(b) — processing necessary for the performance of the contract — combined with Article 9(2)(a) explicit consent where health data is processed for model improvement.

• On-device processing: By default, audio is analysed locally in the browser and never transmitted. Only derived metrics (event counts, duration, stage scores) are stored.
• Session recording (opt-in): Users who opt in to session recording consent explicitly at onboarding. Recordings are encrypted in transit (TLS 1.3) and at rest (AES-256) in Supabase Storage (EU West region, Frankfurt).
• Data retention: Session metrics are retained for 36 months. Audio recordings are retained for 12 months then auto-deleted. Users can export or delete all their data at any time via Account Settings → Data Export.
• Model training: Audio from consented sessions may be used to improve detection models. All training data is anonymised; speaker identity is separated from audio via a one-way hash before any model training batch runs.
• Sub-processors: Supabase (database, storage, EU); Vercel (compute, EU via IAD1); Stripe (payment processing); Resend (transactional email). Full DPA available at flowen.digital/dpa.

A Subject Access Request (SAR) can be submitted by emailing privacy@flowen.digital. Flowen's Data Protection contact is the company director at the registered address.

6. Clinical Safety — DCB0129 Compliance

As a digital health tool used in the context of NHS-funded speech therapy, Flowen is subject to NHS England's DCB0129 Clinical Risk Management Standard. Flowen maintains:

• A Clinical Safety Officer (CSO) appointed from the founding team, with relevant clinical background, responsible for clinical risk management.
• A Clinical Risk Management Plan (CRMP) covering hazard identification, risk assessment, and mitigation for all clinical functions of the platform.
• A Clinical Risk Management File (CRMF) containing all hazard logs, evidence of testing, and incident reports.
• Residual clinical risk classification of Acceptable for all identified hazards at the time of this publication.

7. Accessibility & Funding Pathways

Flowen is committed to making effective stammering support accessible regardless of personal financial means:

• Access to Work (AtW):Adults who stammer in employment can apply to DWP's Access to Work scheme to fund a Flowen Standard subscription. Flowen is a recognised communication support tool. Typical award: 12 months subscription cost, renewed annually. Flowen provides AtW evidence letters on request.
• DSA (Disabled Students' Allowance): Students in higher education can apply for DSA to cover subscription costs. Flowen appears on several university assistive technology lists.
• NHS-funded access: ICBs can commission Flowen as a group service for NHS SLT departments. Flowen offers NHS group pricing with DTAC-aligned evidence packs.
• Public Funds tier: Individuals funded via AtW or NHS receive a complimentary Public Funds subscription tier at no personal cost.

8. Intellectual Property

Flowen's core IP assets are held by Flowen Technologies Ltd and include:

Flowen participates in the UK Patent Box scheme (milestone: first patent granted) and is pursuing R&D Tax Credits (RDEC) for qualifying AI model development expenditure.

9. References

Bloodstein, O., & Ratner, N. B. (2008). A Handbook on Stuttering (6th ed.). Delmar Cengage Learning.

Briley, P. M., & Ellis, C. (2018). The prospective association between stuttering and mental health disorders in a nationally representative sample. Journal of Speech, Language, and Hearing Research, 61(10), 2552–2566.

Craig, A., Blumgart, E., & Tran, Y. (2009). The impact of stuttering on the quality of life in adults who stutter. Journal of Fluency Disorders, 34(2), 61–71.

Cream, A., O'Brian, S., Onslow, M., Packman, A., & Menzies, R. (2019). Self-modelling as a treatment for stuttering: A two-subject multiple baseline randomised trial. Speech, Language and Hearing, 22(2), 109–118.

Euler, H. A., Gudenberg, A. W. V., Jung, K., & Neumann, K. (2014). Computerised treatment of stuttering: Perceptual assessment of speech quality. Folia Phoniatrica et Logopaedica, 66(4–5), 176–187.

Lincoln, M., Packman, A., & Onslow, M. (2006). Altered auditory feedback and the treatment of stuttering: A review. Journal of Fluency Disorders, 31(2), 71–89.

Onslow, M., Packman, A., & Menzies, R. (2017). Biofeedback in stuttering treatment. In Handbook of Evidenced-Based Practice in Communication Disorders. Plural Publishing.

Packman, A., Onslow, M., & Menzies, R. (2014). Novel speech patterns and the management of stuttering. Disability and Rehabilitation, 22(1–2), 65–79.

STAMMA. (2023). Facts and figures about stammering. Stammering Association. https://stamma.org/about-stammering/facts-and-figures

Stuart, A., Kalinowski, J., Rastatter, M. P., Saltuklaroglu, T., & Dayalu, V. (2004). Investigations of the impact of altered auditory feedback in-the-ear devices on the speech of people who stutter. International Journal of Language & Communication Disorders, 39(2), 215–227.

van Borsel, J., Reunes, G., & Van den Bergh, N. (2003). Delayed auditory feedback in the treatment of stuttering: Clients as consumers. International Journal of Language & Communication Disorders, 38(2), 119–129.