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 postVoice AI engineering
How we build and run real-time text-to-speech in production: latency, streaming, GPU serving and integration with voice agent frameworks.
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
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
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