Skip to content

Choose a voice architecture

Supported in ADKPython v2.13.0Java v0.2.0

ADK supports two voice architectures under Runner.run_live():

  • Speech-to-speech: A single native audio model processes speech input, reasoning, and speech output end to end over the Gemini Live API. Best when lowest latency, vocal emotion, or live video input matter most.
  • Cascade: Chains speech-to-text (STT), any text-reasoning LLM, and text-to-speech (TTS) in sequence. Best when frontier text reasoning, modular speech providers, or pre-synthesis text guardrails matter most.

Both architectures use the same ADK runtime. Your session loop, request queue, event stream, tools, callbacks, and workflows stay identical, and switching architectures only changes the model on your agent:

from google.adk.agents import Agent

root_agent = Agent(
    name="voice_assistant",
    model="gemini-live-2.5-flash-native-audio",
    instruction="You are a helpful voice assistant.",
    tools=[...],
)

from google.adk.agents import Agent
from google.adk.live import CascadeLive, GeminiSTT, GeminiTTS

root_agent = Agent(
    name="voice_assistant",
    model=CascadeLive(
        model="gemini-3.8-flash",
        stt=GeminiSTT(),
        tts=GeminiTTS(speech_config="Kore"),
    ),
    instruction="You are a helpful voice assistant.",
    tools=[...],
)

Architecture comparison

flowchart LR
    subgraph S2S["Speech-to-speech"]
        direction LR
        U1["User Audio"] --> M1["Native Audio Live Model"] --> A1["Agent Audio"]
    end
    subgraph CAS["Cascade"]
        direction LR
        U2["User Audio"] --> STT["Speech-to-Text"] --> LLM["Text LLM"] --> TTS["Text-to-Speech"] --> A2["Agent Audio"]
    end
Dimension Speech-to-speech Cascade
End-to-end latency Lowest (single model pass) Moderate (serial stages overlapped via streaming and sentence prefetching)
Reasoning model Native audio Live models only (Supported models) Any text LLM (Gemini Flash, Gemini Pro, or third-party models)
Speech providers Built-in Gemini Live voices Modular (Gemini, ElevenLabs, or custom STT/TTS providers)
Vocal nuance Hears tone and emotion directly; native prosody STT converts speech to text; TTS synthesizes sentence by sentence
Video input Supported (live camera and screen frames) Not supported (audio and text only)
Text guardrails Screens transcriptions alongside audio generation Full text boundary between stages; can inspect or modify text before synthesis
Text-only execution (run_async) Not supported Supported (uses the inner text LLM directly)
SDK support Python, Java Python (Experimental)