--- id: wiechmann26_interspeech title: Can deep learning based voice editing enhance voice quality perception skills in speech therapy students? authors: - Jana Wiechmann - Frederik Rautenberg - Reinhold Haeb-Umbach - Petra Wagner year: 2026 doi: 10.21437/Interspeech.2026-2475 isca_url: https://www.isca-archive.org/interspeech_2026/wiechmann26_interspeech.html pdf_url: https://www.isca-archive.org/interspeech_2026/wiechmann26_interspeech.pdf session: Speech and Language Technologies for Health Applications 1 topics: - speech-enhancement - evaluation - paralinguistics category: health-clinical labels: - generative-model institutions: - Bielefeld University - Paderborn University funding: - Deutsche Forschungsgemeinschaft code: url: "" license: "" open_to_collaboration: false wiki_frontmatter: id: wiechmann26_interspeech category: health-clinical labels: - generative-model institutions: - Bielefeld University - Paderborn University updated: 2026-09-29 confidence: full-paper digest: v2 source: https://doi.org/10.21437/Interspeech.2026-2475 pdf: https://www.isca-archive.org/interspeech_2026/wiechmann26_interspeech.pdf wiki_url: https://interspeech-2026-wiki.vercel.app/papers/wiechmann26_interspeech/ markdown_url: https://interspeech-2026-wiki.vercel.app/papers/wiechmann26_interspeech/markdown.md --- # Can deep learning based voice editing enhance voice quality perception skills in speech therapy students? *Jana Wiechmann, Frederik Rautenberg, Reinhold Haeb-Umbach, Petra Wagner* [PDF](https://www.isca-archive.org/interspeech_2026/wiechmann26_interspeech.pdf) · [ISCA page](https://www.isca-archive.org/interspeech_2026/wiechmann26_interspeech.html) · [DOI](https://doi.org/10.21437/Interspeech.2026-2475) **Category:** `health-clinical` · **Labels:** `generative-model` **TL;DR** — Using deep learning-based speech synthesis to isolate voice characteristics during training significantly improves speech therapy students' perceptual sensitivity and gold-standard agreement compared to natural anchor voices. The synthesis group achieved a post-test sensitivity (d') of 1.03 versus 0.55 for the control group. ## Key contributions - Evaluates a deep learning voice editing system directly in an applied educational setting for speech therapy students. - Demonstrates that isolated synthetic prototypes improve perceptual sensitivity (d') and gold-standard agreement (Cohen's kappa) over natural anchor voices. - Shows that sensitivity to difficult-to-perceive voice qualities like 'rough' can be improved from below-chance to above-chance levels using controllable speech synthesis. - Provides an interactive experimental paradigm pairing a ~10-minute expert explanation with fine-grained multi-strength synthetic voice variations. ## Problem Auditory-perceptual assessment of voice quality is notoriously difficult and unreliable even for experts, posing a steep learning curve for novices like speech therapy students. Natural voices are inherently multidimensional, meaning a single voice simultaneously exhibits multiple overlapping acoustic features like breathiness, roughness, and creakiness, making it impossible for listeners to experience features in isolation. Prior training tools relying on natural anchor voices fail to decouple these dimensions, leading to high variability in listener judgments and low diagnostic accuracy. ## Method The study utilizes a neural text-to-speech (TTS) framework equipped with a manipulation block built on Conditional Continuous Normalizing Flows. This block transforms a global speaker representation by taking the strength levels of seven specific voice qualities as conditioning inputs, enabling the independent modification of breathy, creaky, and rough dimensions while preserving naturalness. Twenty clinical linguistics students participated in a between-subjects design split evenly into a control group (trained with natural anchor voices) and a synthesis group (trained with synthetically generated isolated prototypes). The experimental workflow consisted of a pre-explanation binary rating phase (16 non-pathological German voices from the NSC corpus evaluated for present/absent traits), a ~10-minute interactive expert explanation phase using respective anchors, and a post-explanation rating phase. Statistical evaluation was performed using paired t-tests, independent t-tests, linear mixed-effects models (LMER) with Kenward-Roger degrees of freedom approximation, and ANCOVA controlling for pre-test scores, treating participant as a random factor. ## Experimental setup The experiment evaluated 20 female clinical linguistics students (aged 18-27) using 16 non-pathological German voice samples (8 female, 8 male) from The Nautilus Speaker Characterization Corpus (NSC). Baselines were trained using natural anchor voices and expert imitation, while the treatment group used deep learning-generated synthetic exemplars. Metrics included perceptual sensitivity (d') and agreement with expert gold standard (Cohen's kappa, where expert-expert subset reliability was kappa = 0.85). ## Results The synthesis group exhibited a statistically significant increase in perceptual sensitivity from pre-test to post-test (d' increasing from 0.73 to 1.03, t(9) = -2.70, p = 0.024), whereas the control group showed no significant change (0.49 to 0.55, t(9) = -0.38, p = 0.71). At the post-test timepoint, the synthesis group achieved significantly higher sensitivity (d' = 1.03 vs 0.55, t(15.77) = -3.08, p = 0.007) and gold-standard agreement (kappa = 0.33 vs 0.19, t(15.33) = -2.55, p = 0.022) compared to controls. An ANCOVA controlling for baseline differences confirmed a significant main effect of group (t(17) = 2.66, p = 0.017). In dimensional breakdowns, only the 'rough' voice characteristic showed a significant improvement in d' for the synthesis group (t(8) = 2.67, p = 0.028), while 'breathy' and 'creaky' showed no notable gains due to high pre-test baseline ratings and potential ceiling/plateau effects on non-pathological subtle stimuli. | System / Condition | Pre-test d' (M ± SD) | Post-test d' (M ± SD) | Pre-test Kappa (M ± SD) | Post-test Kappa (M ± SD) | |---|---|---|---|---| | Control Group (Natural) | 0.49 ± 0.50 | 0.55 ± 0.40 | 0.16 ± 0.16 | 0.19 ± 0.14 | | Synthesis Group (Deep Learning) | 0.73 ± 0.41 | 1.03 ± 0.27 | 0.25 ± 0.14 | 0.33 ± 0.09 | ## Limitations The study relies on a relatively small sample size (n = 20 female students from a single university), limiting the statistical power of the mixed-effects models. The interactive nature of the explanation phase prevented blinding of the expert explainer to experimental conditions, introducing potential explainer bias. Furthermore, testing was restricted to non-pathological German speech samples over a short 10-minute training window, leaving long-term retention and generalization to pathological voices unexplored. ## Why read this Speech and ML researchers building controllable voice generation systems should read this paper to see a practical validation of neural voice editing applied to human perceptual training and clinical education. It demonstrates that disentangling acoustic dimensions via normalizing flows directly translates to measurable cognitive improvements in human listeners. ## Code None released (as of this page's `updated` date). If you are an author with a repo, please claim this entry — see CONTRIBUTING.md. ## Applications Speech therapy student training, clinical voice assessment software, computer-aided phonetic education tools, and interactive perceptual training modules for forensic phonetics. ## Institutions / 機構 Bielefeld University, Paderborn University **Funding / 經費:** Deutsche Forschungsgemeinschaft ## Related - [A barrier or a booster? Familiarity effects on Mandarin emotion prosody recognition using AI-powered voice cloning](xu26i_interspeech.md) — shared technique · relatedness 1.9/3 - [Synthetic Speech, Real Signal: Paralinguistic Preservation and Cross-Lingual Augmentation via Voice Cloning](polle26_interspeech.md) — shared technique · relatedness 1.9/3 - [Intonation Perception in Real and Synthetic Speech across Varying Familiarity Levels: A Pilot Study of Equivalence Assessment](zhou26b_interspeech.md) — shared technique · relatedness 1.9/3 - [Synthetic Pathological Speech at Scale: A Flow Matching Approach for Clinical Data Augmentation](koudounas26_interspeech.md) — shared technique · relatedness 1.8/3 - [VoiceQualityGUI: A Tool for Word-Level Voice Quality Modifications](lameris26b_interspeech.md) — shared technique · relatedness 1.7/3 All 950k paper pairs scored by TypeSafe Jev (`scripts/related/`); relatedness 0 = unrelated … 3 = directly comparable.