SpeechPath

SpeechPath AI - Inspiration & Journey

Inspiration

• 16 million children with speech disorders face a therapy access gap
• Speech therapists wanting more time for breakthrough moments, not reteaching basics
• Research proving 30-50% improvement with increased practice frequency yet this gap remains unfilled

What It Does

• Analyzes children's speech using Whisper ASR + acoustic feature extraction
• Generates real-time, clinical-grade feedback (articulation, fluency, clarity scores)
• Creates personalized, game-based practice scenarios with LLM (never the same scenario twice)
• Provides child-friendly feedback with stars, praise, and specific guidance (never says "wrong")
• Gives SLPs complete data: which sounds improved, engagement patterns, practice frequency

How We Built It

• Speech Analysis: Whisper for transcription + librosa for audio feature extraction (pauses, stutters, prolonged sounds)
• Scoring Engine: Articulation (onset clarity + spectral quality + confidence), Fluency (penalty-based), Clarity (SNR analysis)
• LLM Integration: GPT-4 for response evaluation (acoustic metrics → child-friendly feedback) + scenario generation (personalized practice activities)

Challenges We Ran Into

• Translating acoustic metrics into clinical judgments that LLMs understand
• Ensuring LLM feedback is always affirming, never discouraging (prompt engineering intensive)
• Handling audio file uploads reliably with proper error handling
• Confidence score extraction from Whisper's token probabilities (not straightforward)

Accomplishments We're Proud Of

• End-to-end system: from audio analysis to child-friendly feedback in <500ms
• AI-generated scenarios that sound professionally designed, not robotic
• Scalable architecture
• Privacy-first approach: speech analysis never leaves the device

What We Learned

• Confidence scores from Whisper require careful extraction from token probabilities
• Children respond better to game-based scenarios than traditional practice exercises
• LLM prompts need extreme specificity to generate consistently affirming feedback
• Motor learning research validates our core insight: practice gaps are where progress dies
• SLPs want data, not features—they care about which sounds improved and why

Built With

• anything
• fastapi
• lovable
• python
• react