Voice Restoration Research Notebook

Voice Restoration Research Notebook

Research Purpose

To explore how wearable sensors and embedded AI can improve the clarity, naturalness, and usability of electrolarynx devices through real-time signal analysis, intent detection, and adaptive assistance.

Current Reference Technology

The TruTone Plus and similar consumer electrolarynx systems represent the current state of practice:

• Hand-held or neck-worn devices with on/off control
• Fixed fundamental frequency output
• Limited adaptation to user intent or speech context
• User manages device activation, frequency, and amplitude manually

Initial Usability Observations

From literature review and user-centered research context:

• Speech clarity is limited by device positioning and vibration transmission
• User must coordinate device activation with speech intent
• Pitch control requires conscious effort, limiting naturalness
• No acoustic feedback during speech production
• Device discomfort and social stigma remain significant barriers

Human-Centered Research Questions

1. How can we detect speech intent *before* acoustic output?
2. What sensor fusion approaches best capture the speaker's intended prosody and timing?
3. Can embedded AI provide real-time pitch and amplitude assistance that feels natural and reduces cognitive load?
4. What wearable form factors balance functionality, comfort, and social acceptance?
5. How should user feedback loops be designed to preserve autonomy and avoid over-automation?

Sensor and AI Research Areas

Potential Sensor Modalities:

• Contact/bone-conduction microphones (throat, neck, jaw)
• MEMS accelerometers (capturing neck and laryngeal vibration)
• Electromyography (EMG) for intent sensing
• Inertial measurement units (IMU) for device orientation and user gesture

Signal Processing Focus:

• Speech-onset detection from subvocal or minimal-effort signals
• Fundamental frequency estimation from multiple sensors
• Formant analysis for speech clarity enhancement
• Noise rejection in real-world environments

Embedded AI Responsibilities:

• Sensor fusion and feature extraction
• Intent classification and timing prediction
• Adaptive audio synthesis (pitch, amplitude, timing)
• Personalization and user preference learning

Potential Research Phases

Phase 1 (Current): Sensor characterization and baseline signal analysis

• Evaluate each sensor modality independently
• Understand signal quality and information content
• Map relationships between user intent and measurable signals

Phase 2: Multi-modal sensor fusion

• Combine complementary sensors
• Develop feature representations
• Explore signal-to-intent relationships

Phase 3: Embedded AI development

• On-device feature extraction and classification
• Real-time pitch and amplitude prediction
• Personalization and user feedback integration

Phase 4: Prototype integration and user testing

• Integrate AI and sensors into a wearable form factor
• Usability studies with electrolarynx users
• Iterative design refinement

Phase 5: Clinical evaluation and regulatory pathways

• Partner with clinical researchers and audiologists
• Understand appropriate regulatory frameworks
• Design for clinical validation if appropriate

Current Status

Early research and discovery phase. This work is exploratory; no claims of clinical effectiveness or device performance. All findings are preliminary.

Next Steps

• Review sensor datasheets and specifications
• Design initial sensor evaluation experiments
• Establish partnerships with speech-language pathologists and users
• Develop signal analysis and feature-extraction approaches