Benchmarks5 min read
TTS latency benchmark: 50 ms to first audio
Four vendors, both transports, 240 measured calls, zero errors. Gepard returns first audio in 49 to 50 ms over WebSocket, about half the next fastest.
Read the post
Nursultan Bakashov
Co-founder, nineninesix.ai
Co-author of the Gepard technical report. Works on real-time speech models and the infrastructure that serves them, and writes about the parts of text-to-speech that only show up in production.
Benchmarks6 min read
TTS API pricing in 2026: the real cost
Ten realtime voice vendors in one table. ElevenLabs Flash is $50 per million characters, Cartesia $30, ours $5. Here is what that means on your monthly bill.
Read the post
Engineering5 min read
How to fine-tune Gepard (and when not to)
Zero-shot cloning takes minutes and covers most cases. A LoRA fine-tune needs about 1 hour of clean audio, a full one needs 4 GPUs. Where each line falls.
Read the post
Engineering5 min read
Self-hosted TTS or an API: the crossover
The crossover sits near 290M characters a month, about 4,800 hours of audio. Below it the API wins on total cost, and the GPU bill is not the whole story.
Read the post
Product5 min read
Open weights or closed TTS: who wins when
Apache 2.0 grants more than most teams think. 4 decision branches, the licence traps in our own catalogue, and the 300M character mark where self-hosting wins.
Read the post
Engineering5 min read
Why Gepard runs on vLLM, and what it buys
One design rule shaped the whole model: it must run on stock vLLM without patching a single compute kernel. The result is 204x real time on one GPU.
Read the post
Research5 min read
KaniTTS to Gepard: which model to use
Three generations and 28 repos on Hugging Face. Gepard 1.0 is 555.7M parameters under Apache 2.0; KaniTTS weights are not. Which model to actually use.
Read the post
Engineering6 min read
How text to speech works, and how to choose
How modern TTS works, the 5 metrics vendors quote and which of them lie, and 8 questions to ask first. Why 200 ms of synthesis eats a whole latency budget.
Read the post