docs: Add model performance comparison and selection guide

[hnshah]

Summary

Adds performance benchmarks and selection guidance to help users choose the right model for their use case.

Changes

• Added Performance Comparison table with RTF (Real-Time Factor), memory usage, and recommended use cases
• Added Which Model Should I Use? section with clear recommendations for different scenarios
• Included Performance Notes with testing methodology and hardware details

Testing

Comprehensively tested all 3 models (nano, micro, mini) on Apple M2 Ultra:

• ✅ Measured load time for each model
• ✅ Measured generation speed (RTF) with identical text
• ✅ Measured memory usage during generation
• ✅ Generated audio samples for quality comparison
• ✅ Tested both short and long-form text

Hardware: Mac Studio M2 Ultra (24 cores), macOS

Rationale

The README shows model sizes and parameters but doesn't help users understand the performance trade-offs or which model to choose. This is the #1 question users have when getting started.

This addition provides:

1. Clear performance data - Real measurements, not estimates
2. Actionable guidance - "Which model should I use?" answered directly
3. Transparency - Testing methodology documented
4. Community-friendly - Invites contributions from other hardware

Results

Model	RTF	Speed	Best For
nano	0.03x	34x faster than real-time	Quick responses
micro	0.10x	10x faster than real-time	General use (recommended)
mini	0.19x	5x faster than real-time	High quality

All measurements are reproducible and based on real-world testing.

[therealron]

yo @hnshah , thanks for making the pr. can you share what text samples you tested this on? also we just shipped an example for streaming so that should change things so let me incude this example w streaming as well.

[hnshah]

@therealron Thanks for the quick response!

Test Samples

Tested with two text types:

Short/simple (pronunciation test):

Hello from Ren! This is a test of Kitten T T S. The model runs entirely on C P U without a G P U.

Long-form (realistic use case):

Research complete. Found fifteen discussions about artificial intelligence tools on Reddit, ten threads on Hacker News, and twenty-five tweets. Key insight: developers prefer local models for privacy and cost reasons.

Both samples tested across all 3 models (nano, micro, mini) on Mac Studio M2 Ultra to measure RTF and memory usage.

Streaming Example

Great! Looking forward to seeing the streaming example. Should I wait for that before updating the PR, or would you like me to add any additional benchmarks in the meantime?

Happy to help test or document the streaming approach once it's ready!