Project Update: GhostCall — Living Dialects
GhostCall: Living Dialects has evolved significantly since its initial concept. Below is a running log of how the project has grown, what changed, and why those changes matter.
Evolution Log
Phase 1 — Concept Prototype
Status: Early experiment
What existed:
- A basic real-time voice interaction prototype
- Single-language conversational flow
- Focused on low-latency speech generation
Key insight:
Fluent speech ≠ being understood. Correct grammar alone didn’t produce natural conversations.
Phase 2 — Shift to Dialect Intelligence
What changed:
- Reframed the project from “voice assistant” to dialect-aware Voice AI
- Introduced the idea of learning through misunderstanding
- Removed text UI and translation features entirely
Why:
Text explanations were interfering with how people naturally adapt their speech.
Phase 3 — Amazon Nova Integration
Major update:
- Integrated 0 as the core reasoning layer
- Replaced rule-based checks with Nova-powered reasoning
- Added pre-speech validation: the system now decides whether it should speak at all
Impact:
This enabled intentional pauses, hesitation, and silence as valid responses.
Phase 4 — Living Dialects Framework
New features introduced:
- Dialect-specific conversational boundaries
- Scenario-based sessions (e.g., market negotiation, casual peer talk)
- Dialect Integrity Agent to prevent accent normalization and code-switching
Result:
Speech that sounds “correct” but feels socially wrong is no longer accepted.
Phase 5 — Multilingual & Dialect Expansion
Added dialect demos:
- Nigerian Pidgin (Urban informal conversation)
- Andean Spanish (Rural market negotiation)
- Indian English (College campus discussion)
Each dialect is:
- Voice-only
- Context-locked
- Session-isolated
No English fallback. No subtitles.
Phase 6 — Demo & Evaluation Readiness
What’s new:
- Judge-only testing flow
- Ephemeral, single-session memory
- Synthetic dialect datasets for safety and cultural respect
- Tight 3-minute demo experience designed around sound, not screens
Design decision:
If the experience can’t be understood by listening, it doesn’t belong in the demo.
Current State
- Voice-only interaction
- Real-time dialect reasoning
- Intentional misunderstanding as a teaching mechanism
- Privacy-first, non-persistent sessions
What’s Next
Planned explorations include:
- Community-curated dialect scenarios
- Accessibility-focused voice pacing modes
- Evaluating GhostCall for cultural preservation and language documentation use cases
GhostCall continues to explore a simple question:
What if Voice AI learned when **not* to understand you?*
More updates coming soon
Log in or sign up for Devpost to join the conversation.