SpeechifyAI

A few weeks ago I sat down to answer what sounded like a simple question: which text-to-speech model should we actually use? I figured I'd check a leaderboard, pick the top one, and move on. That's not how it went. By the time I'd worked through 19 models, the "winner" had changed twice. Cartesia's Sonic 3.5 took the number one spot on May 22. Nine days later, Google's Gemini 3.1 Flash TTS had taken it back. The ground was moving while I was standing on it. So I stopped hunting for the best model and started asking a better question: best at what? Once I framed it that way, the picture got clear. If you're building a live voice agent, latency is everything, and Sonic 3.5 answers in about 82 milliseconds. If you're producing long narration, Gemini gives you 200+ audio tags and real control, but it can't stream and caps at 32k tokens, so it's built for reading, not conversation. If you want to self-host, Fish Audio S2 Pro is the strongest open-weight model out there, right up until you read the license and realize commercial use costs extra. And if budget is the constraint, Kokoro runs on a CPU for around 65 cents per million characters. None of them is "the best." Each one wins a specific brief. That's the part the leaderboards leave out. A number one ranking answers a question most teams aren't actually asking. The real work is naming the one thing you can't compromise on, then testing two models on your own text. I wrote up all 19, with the trade-offs, pricing, and licensing, in one place. Here is our take: marktechpost.com/2026/05/30/… @inworld @cartesia @OpenAI @GoogleDeepMind @Microsoft @xai @MistralDevs @MistralAI @ElevenLabs @DeepgramAI @hume_ai @SpeechifyAI @Alibaba_Qwen @MiniMax_AI