Question 1

What does "swap providers per agent" actually mean?

Accepted Answer

Each agent definition has a config flag for LLM, STT, TTS, and avatar. Change the value and redeploy — that's the entire swap. No code changes, no test rewrites. Useful for cost-routing (Haiku for triage, Opus for complex) or for falling back when a provider has an outage.

Question 2

Can I use my own API key with each provider?

Accepted Answer

Yes, on Growth and Enterprise. Bring your own OpenAI, Anthropic, Gemini, ElevenLabs, or Cartesia keys. Saaya routes through your account, so usage and billing live where you already manage them.

Question 3

How do you keep latency low across so many vendors?

Accepted Answer

A real-time pipeline with per-leg streaming (STT chunks → LLM tokens → TTS audio) and aggressive parallelism. The Session Viewer shows latency per leg so you can see exactly where time is being spent and switch the slow provider out.

Question 4

What happens if a provider has an outage?

Accepted Answer

Saaya supports per-agent fallback chains. Define an ordered list of providers; we fail over within seconds and log the event. Most teams set a primary + secondary for production-critical agents.

Question 5

Are non-English voices any good?

Accepted Answer

Indic-language coverage via Sarvam is native-quality, including code-switching (Hinglish). European and East-Asian languages run on ElevenLabs, Cartesia, and OpenAI voices. We benchmark new voices monthly — the list updates as the underlying models do.

Question 6

Can we add a provider you don't support yet?

Accepted Answer

Yes — talk to us. The provider abstraction is small (auth, request, stream); most new providers are a 1–2 day plug-in. Enterprise customers can also bring proprietary models via OpenAI-compatible endpoints.

Question 7

How do you handle pricing differences between providers?

Accepted Answer

Per-session cost breakdown is in the Session Viewer — you see the LLM, STT, TTS, and avatar costs separately. Set per-agent budget caps and cost-routing rules to keep cheap calls cheap.

Best-in-class models. Zero vendor lock-in.

The model wars never end. Your platform shouldn’t care.

Switch in seconds

Cost-route per call

Failover, not lock-in

Five layers. Pick the best for each.

LLMs

Speech-to-text

Voices (TTS)

Avatars

Telephony & messaging

Frequently Asked Questions.

Best models. No lock-in. Ship today.