Ultravox.ai
One of the few voice stacks where the model itself is the product rather than glue over someone else's transcriber, and the open weights on Hugging Face mean you are not locked into the hosted service. Skipping the speech-to-text hop is the whole point — latency and prosody both benefit. It is a builder's tool, not a no-code agent designer, so budget engineering time as well as per-minute cost.
Pros
- Model takes audio directly, removing the transcription hop from the response path
- Open-weight models published on Hugging Face allow self-hosting
- REST API plus web and mobile SDKs, and SIP for phone calls
- RAG corpora and custom voices are part of the platform, not add-on services
- Pay as you go has no monthly floor, so a prototype costs only what it talks
Cons
- Free tier is limited to 5 concurrent calls, so load testing needs a paid plan
- Custom voices and RAG corpora are capped even on the $100 Pro tier
- Threads bill separately per token on top of per-minute conversation cost
- Requires real engineering; there is no visual flow builder