A useful way to view modern in-context TTS is through four overlapping paradigms: 1. Autoregressive discrete-token systems Examples: VALL-E, Spark-TTS, Qwen3-TTS. They integrate naturally with language modeling and support causal, low-latency generation, but quantization may discard fine acoustic detail and decoding cost grows with token rate. 2. Non-autoregressive continuous systems Examples: Voicebox, E2 TTS, F5-TTS. They use diffusion or flow matching to generate continuous acoustic representations in parallel, often achieving high fidelity, but iterative utterance-level sampling can make streaming more difficult. 3. Hybrid systems Examples: Seed-TTS and the CosyVoice family. They combine autoregressive semantic planning with continuous acoustic rendering, separating linguistic organization from detailed speech generation. The remaining bottleneck is often the discrete interface between the two stages. 4. Continuous-autoregressive systems Examples: DiTAR, Dots.TTS, VoxCPM2. They generate continuous latent patches autoregressively without an external speech tokenizer, avoiding quantization loss, but high-dimensional prediction, local iterative sampling, and error propagation can affect stability and long-form consistency. These approaches offer complementary trade-offs across fidelity, latency, streaming, controllability, and robustness. There is still no universally dominant architecture. The harder challenge is building a production-grade system that delivers all of these strengths simultaneously: preserve linguistic content and speaker identity; generate clean, natural, and expressive speech; follow flexible control instructions; support diverse languages, accents, and dialects; stream with low latency; remain stable with noisy, reverberant, or bandwidth-limited prompts. Standard short-form, clean-speech benchmarks capture only part of this problem. They can easily hide failures in multilingual synthesis, dialect coverage, long-form stability, streaming behavior, and adverse acoustic conditions. The frontier of TTS is no longer just intelligibility or naturalness in ideal conditions. It is whether one system can remain high-quality, controllable, multilingual, low-latency, and robust in the real world.