Work
Hip-Hop Voice AI - Cultural Voice Preservation
Overview
Blacksky built a voice-cloned conversational AI that preserves the voice, personality, and cultural knowledge of a legendary hip-hop figure. The project combines RAG (Retrieval-Augmented Generation) with voice synthesis to create an interactive experience that honors hip-hop cultural heritage. This proof of concept demonstrates Blacksky's full LLM stack capabilities - from data pipeline to voice output.
The Challenge
When influential cultural figures pass away, decades of knowledge, perspective, and lived experience go with them. Hip-hop has lost many of its pioneering voices - attorneys who represented artists, journalists who broke stories, podcast hosts who interviewed legends. Traditional archives preserve recordings, but they're passive. The question Blacksky asked: What if AI could create an interactive archive - a way for future generations to engage with these perspectives directly?
The goal wasn't to replace or imitate anyone, but to build technology that could preserve cultural voices as living, conversational knowledge bases.
Blacksky's Approach
Blacksky designed a three-node architecture that separates concerns cleanly:
Brain Node - Handles knowledge retrieval and response generation using Llama 3 with RAG. The knowledge base was built from podcast transcriptions, creating a searchable archive of conversations, interviews, and commentary.
Bridge Node - Custom middleware that orchestrates requests, injects personality characteristics, and ensures responses stay true to the subject's speech patterns, humor, and cultural references.
Voice Node - Coqui XTTS v2 voice cloning that synthesizes responses in the subject's actual voice, trained on clean audio samples from original recordings.
A critical architectural decision was 100% local deployment. Everything runs on a MacBook Pro M3 - no cloud APIs, no data leaving the machine, no ongoing costs. This proves the technology is accessible, not just for enterprises with cloud budgets.
Key Accomplishments
- Complete data pipeline: Podcast audio ingested, transcribed via OpenAI Whisper, chunked semantically, embedded as vectors, and stored in LanceDB for retrieval
- Local LLM deployment: Llama 3 running via Ollama on Apple Silicon, proving enterprise-grade AI can run on consumer hardware
- Voice cloning: Coqui XTTS v2 trained on audio samples to reproduce the subject's voice with natural speech patterns
- Hybrid CPU/GPU strategy: Solved Apple Silicon MPS compatibility issues by running embeddings on CPU and inference on GPU
- Personality tuning: System prompts carefully calibrated to capture speech patterns and cultural references without caricature
Technologies
| Component | Technology | Why This Choice | |-----------|------------|-----------------| | LLM | Llama 3 via Ollama | Open source, runs locally on Apple Silicon | | RAG Framework | AnythingLLM | Flexible RAG orchestration | | Vector Database | LanceDB | Lightweight, no server required, fast | | Voice Synthesis | Coqui XTTS v2 | Best open-source voice cloning available | | Transcription | OpenAI Whisper | Industry-leading accuracy for speech-to-text | | Backend | Python, FastAPI | Fast API development, async support |
Results & Impact
The proof of concept is fully functional - a user can have a voice conversation with an AI that speaks in the subject's voice and draws on their actual recorded knowledge. This demonstrates:
- Full LLM stack mastery: Data pipeline, embeddings, RAG, inference, voice synthesis - end to end
- Edge deployment: Enterprise AI capabilities on a laptop, no cloud required
- Cultural technology: AI serving communities often left out of tech innovation
Why This Project Matters
This project showcases Blacksky's depth across the entire AI stack. Most consultancies can call an API. Blacksky built every layer:
- Data Engineering - Audio ingestion, transcription, chunking, embedding
- LLM Infrastructure - Local deployment, RAG configuration, prompt engineering
- Voice AI - Model training, synthesis, quality optimization
- System Architecture - Three-node design, middleware orchestration
It also reflects Blacksky's commitment to non-European narratives in technology - using AI as a tool for cultural archiving and preservation, not just commercial applications.
For prospects this demonstrates: If you need AI that goes beyond chatbots - custom pipelines, voice interfaces, edge deployment, or culturally-specific applications - Blacksky has built it from scratch.
Project Type: Internal R&D / Proof of Concept Timeline: 2024-2025 Deployment: Local (Apple Silicon M3) Status: Complete - functional proof of concept