STT → LLM → TTS

Cascading

Speech-to-text, a language model and text-to-speech, each streaming into the next. Choose the model for every stage, and see where the time goes on every call.

Speech-to-text
5 models, 2 providers
Language models
20+ models, 3 providers
Voices
41 voices, 3 providers
  1. Listening
  2. Thinking
  3. Speaking
How a turn flows

Nothing waits for a full sentence

One question from a caller, followed through every stage of the pipeline.

Callerspeaks
Speech-to-textDeepgram or Grok

Can I move my appointment to Thursday? 

Turn detectionend of turn
done
Language modelGPT, Claude or Gemini

Sure. Thursday at 9:30 is open, want it? 

Text-to-speechagent speaks
Speech-to-text transcribes while the caller is still talking, and the voice starts on the first words of the answer. Each stage begins before the one above it has finished. Timing shown is illustrative.
Live demo

Talk to a cascading agent

This agent runs the cascading pipeline. Interrupt it, change the subject, and watch the transcript build next to it.

Stages

Pick every stage

Every stage is a separate component, so you can mix providers inside one agent.

  • Speech-to-text

    Streaming transcription, so the model has most of the sentence before the caller finishes. Language is set per agent.

    Deepgram
    • Flux (English)
    • Flux (Multilingual)
    • Nova 3
    • Nova 2
    xAI
    • Grok STT
  • Language model

    Pick the model that fits the call, from small and fast to most capable. Temperature runs from 0 to 2 per agent.

    OpenAI
    • GPT 4.1
    • 4.1 Mini
    • 4.1 Nano
    • GPT 4o
    • 4o Mini
    • GPT 5
    • 5 Mini
    • 5 Nano
    • 5.1
    • 5.2
    • 5.4
    • 5.4 Mini
    • 5.4 Nano
    • 5.5
    • 5.6 Sol
    • 5.6 Terra
    • 5.6 Luna
    Anthropic
    • Claude Sonnet 5
    • Sonnet 4.6
    • Opus 4.8
    • Haiku 4.5
    Google
    • Gemini 3.6 Flash
    • 3.5 Flash-Lite
  • Text-to-speech

    41 voices across three providers, with accent, gender and speed options. Listen to a sample before you choose.

    OpenAI · 10 voices
    • gpt-4o-mini-tts
    ElevenLabs · 26 voices
    • v3
    • v3 Conversational
    • Flash 2.5
    xAI · 5 voices
    • Grok TTS
  • Turn detection

    Decides when the caller has finished rather than paused, so the agent neither cuts in nor leaves dead air.

    • 23 languages
    • ~12 ms inference
    • Silence threshold per agent
  • Latency on every call

    Each call records the model’s time to first token, the voice’s time to first byte and the end-to-end time, so a slow turn shows you where it lost time.

Which engine fits?

Same tools, knowledge and phone numbers on all three. The engine is chosen per agent.

CascadingSTT → LLM → TTS
How it works
Speech-to-text, then a language model, then text-to-speech
Response feel
Fast, with every stage streaming
Control
Highest. Swap STT, model and voice independently
What you can inspect
Transcript, model and voice latency, end to end
Reach for it when
You need a specific voice, a specific model, or strict tool use
Speech-to-speechAudio in, audio out
How it works
One realtime model takes audio in and returns audio
Response feel
Fastest to first word; no transcription hop
Control
Choose the model and its voice; one provider per call
What you can inspect
Transcript and end-to-end latency
Reach for it when
Speed and natural tone matter more than per-stage control
Full duplexListens while it talks
How it works
Keeps listening while it speaks, so callers can overlap and interrupt
Response feel
Most like a person; overlaps, backchannels, barge-in
Control
Tune interruption sensitivity and listening cues
What you can inspect
Full transcript and call detail
Reach for it when
Callers talk over each other and the agent must keep up

Tell us what you are building

Tell us what you are making. We will walk you through the engines, the API, and what it takes to run it in production.