Sai Raina, Suraj Shivaramakrishnan

SpeechPath AI: Intelligent Speech Therapy Practice Between Sessions

• 16 million children in the US have communication disorders
• Most need speech therapy 3x per week but only receive 1x per week
• Research shows 30-50% improvement in outcomes with increased practice frequency
• The gap between sessions is where progress stalls—children have nothing meaningful to do at home

SpeechPath AI is an AI-powered practice companion that fills the therapy gap:

• Interactive, game-based scenarios designed by speech-language pathologists
• Real-time AI feedback using speech recognition and motor learning principles
• Child-friendly evaluation with stars, encouragement, and specific guidance
• SLP dashboard showing practice patterns, accuracy trends, and engagement
• Runs on any smartphone (iOS/Android)—no special hardware needed

Not a replacement for therapists. A multiplier for their impact.

1. Child opens app, selects a fun practice scenario
2. App presents story-based practice (e.g., "Help order pizza, practice /s/ sounds")
3. Child speaks into device
4. AI analyzes speech in real-time—articulation, clarity, fluency
5. Child gets instant, affirming feedback with stars and encouragement
6. Therapist sees the data—what sounds mastered, which need work

Speech Analysis

• Whisper (OpenAI) for speech-to-text transcription
• Audio feature extraction: pauses, stuttering indicators, prolonged sounds
• Articulation scoring: onset clarity + spectral quality + confidence
• Fluency scoring: penalty for pauses, stutters, prolonged sounds
• Clarity scoring: signal-to-noise ratio (SNR) analysis

Response Evaluation (LLM-based)

• Uses GPT-4 to generate child-friendly feedback
• Converts audio metrics into clinical judgments
• Ensures affirming, specific, actionable guidance
• Never uses negative language (never "wrong", always "try again")

Scenario Generation (LLM-based)

• Generates personalized practice activities based on:
  • Target sound/goal
  • Child age and interests
  • Disorder type (articulation, fluency, apraxia, language)
  • Difficulty level
• Embeds target sounds in meaningful contexts (not isolated drills)
• Creates multi-prompt scenarios with visual supports

Frontend (React Native)

• Mobile app for iOS/Android
• Game-based UI for children
• SLP dashboard preview

Backend (FastAPI)

• RESTful API for all operations
• Audio processing pipeline
• LLM integration (OpenAI GPT-4)
• Session management
• Progress tracking

Core Services

• Speech analysis (local inference for privacy)
• Response evaluation (LLM-based)
• Scenario generation (LLM-based)
• Practice session management
• Progress reporting

• Whisper ASR: Transcribe speech with confidence scores
• Feature Extraction: Detect pauses, stuttering, prolonged sounds
• Scoring: Articulation (0-100), Fluency (0-100), Clarity (0-100)
• Output: Structured metrics for LLM evaluation

• Input: Audio metrics + child age + target sound + disorder type
• Process: GPT-4 generates clinical judgment + child-friendly feedback
• Output: Stars (1-3), praise, specific guidance, encouragement
• Example: 85%+ accuracy → ⭐⭐⭐ "Awesome job! Your /s/ was super clear!"

• Input: Child age + target sound + interests + disorder type + difficulty
• Process: GPT-4 creates engaging practice activities
• Output: Story context + 3-5 practice prompts with phonetic focus + visual supports
• Example: "Pizza Party" scenario for /s/ sound with natural contexts ("Please, may I have a slice?")

• Motor Learning: More practice = brain rewires faster (30-50% improvement per research)
• Engagement: Game-based play keeps kids motivated to practice
• Clinical Validation: Scenarios grounded in speech therapy evidence-based practices
• Scalability: LLM-generated scenarios adapt infinitely without new hardware

[]

SpeechPath AI: Filling the practice gap between sessions. Accessible. Scalable. Evidence-backed.