AI TTS Developer Needed for Human-Like Multi-Speaker Podcast Generator

I need a developer to help me build a simple private AI text-to-speech tool for my English-learning company. The main goal is very simple: We paste a script containing a conversation between two or more speakers, choose the voices, click Generate, and get high-quality human-like audio. The most important thing is the quality of the voices and conversation. I want the result to sound natural and human, close to the best AI voice tools available today. We will use this heavily for podcasts, stories, repetitive English practice, and other long-form content, so the generation cost needs to be low enough for large-scale use. What I want: - Natural, human-like American English voices - Support for at least 2–4 speakers - Ability to create or use our own consistent voices - Long-form audio, including 30–60 minute content - Simple controls for speaking speed and pauses - A simple private website/interface - Reliable generation - Low cost per generated hour - Good audio quality even for long conversations I am open to any technical approach. You can suggest: - Open-source models - Self-hosted models - APIs - Cloud GPU solutions - A combination of different technologies I do not want to tell you which model or architecture to use. I want you to recommend the best solution based on quality, reliability, cost, and ease of use. Before I hire someone for the full project, I want to hear a short sample of the quality you believe you can achieve. Please answer: 1. What solution would you recommend? 2. How human-like can the final result be? 3. Can we create 2–4 voices that remain consistent? 4. Approximately how much would generating one hour of audio cost? 5. Can it handle 30–60 minute conversations reliably? 6. How much work would it take to build a simple working version? 7. What would you estimate the total development cost to be? 8. Have you built anything similar before? The first goal is to prove the voice quality and cost. If the result is good, we can continue building the full internal tool.