Empathy Engine

Approach 1 — Implemented (This Demo)

Uses the Inworld TTS API with dynamic vocal parameter modulation driven by a HuggingFace emotion classifier:

• Emotion Detection — j-hartmann/emotion-english-distilroberta-base — 7 classes: joy, surprise, anger, disgust, fear, sadness, neutral
• Intensity Scaling — Model confidence (0–1) linearly scales vocal params — "I'm okay" sounds different from "THIS IS AMAZING"
• Vocal Modulation — speakingRate (speed) + temperature (expressiveness) — both dynamically computed per emotion
• Emotion Mapping — Each emotion has a defined base rate + temperature, scaled by intensity score
• TTS Engine — Inworld TTS inworld-tts-1.5-max, voice: Clive, output: .mp3

Approach 2 — Theoretical (Research-Grade)

Architecturally superior — validated in research, not implemented here due to time and compute constraints:

• Model — F5-TTS with conditional flow matching, learns the full distribution of human speech prosody
• Emotion Transfer — Zero-shot emotion cloning from reference audio — the model hears the target emotion and replicates it
• Why it's better — Flow matching captures micro-variations in pitch, rhythm, and timbre that API parameters cannot replicate
• Papers — F5-TTS: A Fairytale for Flow-matching-based TTS (Chen et al., 2024) · Voicebox (Meta AI, 2023)

This approach remains theoretical in this submission but is well-validated in published research, and would be the production-grade choice given sufficient time and GPU resources.