Pilot of Immersive VoiceSpace VR (N=17, vocally healthy plus people with dysphonia) - participants scaled loudness and pitch across graded virtual restaurant conditions

Background

Voice change is a motor-learning problem, not just a knowledge problem. Behavioral voice therapy is effective for many voice conditions, but gains in the clinic often fail to carry over to everyday communication. The motor-learning literature is clear about why: durable change depends on practicing under conditions that resemble the target context, not just performing the behavior in a structured session. The Specificity of Learning Principle, Transfer-Appropriate Processing, and Encoding Specificity all converge on the same point - when the sensory and contextual demands of practice match the demands of real use, transfer is stronger.

Real-world voice use happens under layered demands: communicative intent, listener distance, social-emotional pressure, room size, background acoustics, and visual-spatial cues that signal how much voice is needed before a person even speaks. Conventional clinic rooms intentionally minimize these variables, which serves initial acquisition but underrepresents the very cues that learning theory says generalization depends on.

Immersive virtual reality offers a controlled way to put those cues back in. Daşdöğen’s 2023 multisensory study (in this Hub) established that visual and audiovisual VR cues drive measurable vocal adaptations in vocally healthy adults, beyond what acoustic simulation alone produces. The 2026 trained-singers study (also in this Hub) extended that to compare expert and untrained speakers. The present study takes the next step: does the same effect hold up in a clinical voice population, and is a custom voice-responsive VR platform feasible to use in that population.

What the researchers did

A within-subjects pilot at Mount Sinai with 17 adults: 10 vocally healthy speakers recruited from otolaryngology clinic and hospital staff, and 7 people with dysphonia recruited during routine voice-assessment visits (diagnoses included presbyphonia, vocal fold polyp, vocal fold paresis, muscle tension dysphonia, and gender-affirming voice care).

The intervention was Immersive VoiceSpace (IVS) - a custom VR platform developed by the sole author. IVS rendered a lightly populated virtual restaurant on an Oculus Quest 3 headset. A waiter non-player character served as the listener target. The waiter responded in real time to the participant’s voice: if voice intensity met a preset threshold, the waiter approached and stayed in a listening posture; if it dropped below threshold for longer than a set timeout, the waiter walked away.

Three parameters were graded across conditions:

• Listener distance - 5 m (Normal), 10 m (Effortful), 15 m (Calling)
• Voice-activation threshold - +3 dB, +5 dB, +10 dB above each participant’s own baseline-condition SPL
• Walkaway timeout - 5 s, 10 s, 20 s

The speech task across all four conditions was the same: “Order a drink, an appetizer, an entrée, and a dessert.” The Baseline condition was performed with a research-team member acting as the listener in the clinical room at ~2 m. The three IVS conditions were performed in the virtual restaurant in randomized order.

To isolate visual-spatial effects, the restaurant’s ambient audio (background conversation and utensils, which IVS can play) was muted across all experimental conditions. Acoustic recording was made through an AKG C520 head-mounted condenser microphone at 7 cm from the mouth, calibrated to a 30 cm reference, captured at 44.1 kHz / 16 bit via Computerized Speech Lab (CSL).

Outcomes: sound pressure level (SPL, dB) and mean speaking fundamental frequency (mean f0, Hz), each extracted from CSL and analyzed in separate linear mixed-effects models with a random intercept for subject. Fixed effects were Group (Typical, Atypical) and Task Condition (Baseline, Normal, Effortful, Calling). The Group × Task Condition interaction was retained for mean f0 (significant) and dropped from the final SPL model (non-significant). Fixed effects were evaluated with Type III sums of squares and Kenward-Roger approximated degrees of freedom; pairwise contrasts used estimated marginal means with Tukey correction.

A 5-point Likert questionnaire (author-developed, not yet validated) captured four domains after the session: Usability and Interaction, Immersion and Realism, Engagement and Perceived Benefit, Comfort and Safety. Domain scores were averaged; an overall feasibility index was the mean of the four domains. Open-text feedback was reviewed descriptively.

What they found

Sound pressure level. A significant main effect of IVS Level: F(3, 48) = 33.94, p < 0.001. Relative to Baseline, SPL increased by 3.83 dB at Normal, 7.41 dB at Effortful, and 9.04 dB at Calling (all p < 0.001). Normal-to-Effortful and Normal-to-Calling pairwise contrasts were significant; the 1.63 dB step from Effortful to Calling was not (p = 0.450), suggesting a ceiling-like pattern at the highest demand level. The Group main effect was also significant: people with dysphonia produced about 6.88 dB lower SPL on average than vocally healthy speakers. The Group × Level interaction was non-significant and was therefore dropped from the final SPL model - both groups raised loudness in parallel as task demands escalated.

Mean speaking f0. A significant main effect of IVS Level: F(3, 45) = 17.63, p < 0.001. Stepwise increases relative to Baseline (intercept ≈ 201.8 Hz for the typical group) of about 36 Hz at Normal (p = 0.008), 66.6 Hz at Effortful (p < 0.001), and 103.9 Hz at Calling (p < 0.001). The Group main effect was significant, but the Level × Group interaction was also significant: F(3, 45) = 3.94, p = 0.014. Decomposing the interaction: at Baseline the groups did not differ in mean f0 (p = 0.102); at Normal the difference approached but did not reach significance (p = 0.055); at Effortful (p = 0.003) and Calling (p < 0.001) the gap was significant and grew with task demand. The dysphonia group raised pitch with task demands, but to a smaller extent than the vocally healthy group.

Feasibility. Domain scores (out of 5): Usability and Interaction 3.9 (moderate-good), Immersion and Realism 3.4 (moderate, the lowest domain), Engagement and Perceived Benefit 4.0 (good), Comfort and Safety 4.5 (excellent). Overall feasibility index 4.0 (good). No adverse events, no cybersickness, no technical interruptions across the protocol. Average parameter-reconfiguration time between trials was about 2 minutes. Total session length was about 20 minutes per participant.

Qualitative feedback. Participants described the experience as “fun,” “like a video game,” and “a realistic way to practice voice use.” They highlighted the live, responsive behavior of the waiter as the most engaging element. The most consistent piece of negative feedback was the limited interactional behavior of the waiter - participants wanted verbal responses, facial expressions, and gestures during listening turns to make the interaction feel more natural.

Why this matters

For the Evidence Hub, three things are important about this paper:

• First published clinical-population use of a custom voice-responsive VR platform. Prior immersive-VR voice work (including Daşdöğen 2023 and Daşdöğen 2026 trained-singers) was largely in vocally healthy adults. This study extends to people with dysphonia, including diagnostically diverse cases.
• Direct evidence that voice-responsive avatar feedback can elicit graded vocal scaling without explicit clinician cueing. This is the closest published demonstration of a generalization-and-transfer mechanism for voice therapy: the participant adjusts vocal output to functional environmental demands, in real time, in response to non-verbal contextual feedback.
• Comparable behavioral pattern across vocally healthy and dysphonia groups for loudness, with constrained pitch flexibility in the dysphonia group. The SPL finding suggests the contextual mechanism is intact in voice-disordered speakers; the f0 finding is consistent with the wider voice literature on reduced phonatory flexibility in disordered phonation.

For Therapy withVR specifically: this work tested IVS, not Therapy withVR. The broader principle it supports (graded visual-spatial demands elicit functional vocal adaptation) is consistent with the rationale clinicians already use when choosing scenes in Therapy withVR for voice work. Direct equivalence of the avatar-threshold trigger mechanism between platforms has not been studied.

Limitations

The paper is explicit about what this trial does and does not establish:

• Sample size is small (N = 17; atypical n = 7). Subgroup analysis by voice diagnosis is not feasible at this N.
• Single session only. The central claim of the IVS theoretical frame is improved transfer over learning sessions, which this design cannot test.
• No control or comparator condition. There is no imagery-based control, no alternative-treatment comparator, and no waitlist arm. Observed effects across IVS levels are consistent with the visual-spatial manipulation but cannot be cleanly separated from VR-exposure or novelty effects.
• Baseline collected outside the headset. The Baseline-to-Normal comparison confounds task demand with the act of putting on the headset and entering a virtual environment for the first time.
• Audio was muted. Restaurant ambient sound (which IVS can play) was deliberately silenced to isolate visual-spatial effects. This is a clean experimental choice but constrains ecological validity - real restaurants are noisy, and noise is a known driver of vocal adjustment.
• Single virtual context. Only one scene (the restaurant) was tested. The clinical roadmap requires demonstrating the same pattern across multiple contexts (clinic, classroom, workplace, performance, medical settings).
• Feasibility questionnaire is author-developed and not validated. Open feedback is informative but should be treated as descriptive rather than psychometric.
• Significant conflict of interest. The sole author is the inventor of IVS, the holder of a US patent application on the technology, and the sole investigator on this study. There is no inter-rater reliability work, no co-investigator quality check, and no independent replication.
• Avatar interactional limits. Participant feedback flagged the lack of verbal and gestural avatar response as a constraint on perceived realism. This is a development priority for future versions and is also a meaningful threat to interpreting Immersion and Realism scores in the current pilot.

How this fits with the wider Evidence Hub

This study is part of a growing thread of immersive-VR voice work centered on Mount Sinai / Daşdöğen and adjacent voice labs:

• Daşdöğen et al. 2023 (Journal of Voice) - the foundational realism-and-validity work in 31 vocally healthy adults across 18 sensory-input conditions. Established that visual and audiovisual VR cues, not just acoustic cues, change vocal output.
• Daşdöğen and Hitchcock 2026 (Journal of Voice) - trained-singer vs untrained-speaker study using the Rooms module of Therapy withVR. Showed virtual distance cues drive vocal scaling differently in trained versus untrained voices.
• Hoff 2026 (Journal of Voice) - VR-based brief meditation before voice therapy. Different mechanism (state-anxiety regulation rather than direct vocal cueing) but the same direction of travel for VR adoption in voice clinics.
• Leyns et al. 2025 (Journal of Voice) - RCT of VR-based gender-affirming voice training using Therapy withVR. Directly relevant given that IVS is reportedly developing gender-affirming voice modules per Mount Sinai institutional reporting.

The broader landscape: voice-VR is moving from “does the simulation feel real enough to change behavior” (mostly answered: yes) to “does practicing in the simulation transfer to real-world voice use” (largely unanswered, pending multisession longitudinal work). This study sits at the boundary - feasibility and immediate behavioral signal are established for a custom voice-responsive platform; transfer is the next test.

Note on the Immersive VoiceSpace platform. IVS is distinct from Therapy withVR. It is a single-scene, voice-threshold-responsive system invented and patented by the study author. The Mount Sinai institutional report (May 2026, “Hypophonia”) describes ongoing work extending IVS to people with Parkinson’s disease hypophonia, with planned modules for vocal feminization and additional contexts. The IP status of IVS could not be independently verified at the time of this review (see funding/COI field).