messagingintermediate

Voice assistant

Pairs Whisper STT with OpenAI TTS. The user speaks, the agent answers in voice. Round-trip stays below 800ms with the right provider mix. Good for hands-free coding, accessibility, voice copilots.

lines 24cost ~$0.02 per minute of conversationinstall hub://digitorn/voice-assistant

Install in one command

digitorn install hub://digitorn/voice-assistant

What you get

Round-trip under 800ms on a good network
Streaming TTS, the user hears the start while the agent thinks
Conversational answer length (under 30s spoken)
Optional barge-in for natural interrupts

Credentials this template uses

OpenAI Anthropic Claude

Click each one for the credential setup command, scope guidance, and provider details.

The full YAML, ready to fork

app.yaml

1app:2  app_id: voice-assistant3  name: "Voice assistant"4  version: "1.0.0"5  category: "communication"67execution:8  mode: conversation9  entry_agent: voice1011modules:12  voice:13    config:14      stt: { provider: whisper, model: whisper-1 }15      tts: { provider: openai, voice: nova, streaming: true }1617agents:18  - id: voice19    modules: [voice]20    brain:21      provider: anthropic22      model: claude-haiku-4-523      credential:24        ref: anthropic_main25        scope: per_user26        provider: anthropic27    system_prompt: |28      Speak conversationally. Keep answers under 30 seconds.29      No bullet lists, no markdown, only spoken language.

How it works

voice module wires STT + TTS

The voice module wraps a transcription provider and a TTS provider behind one declarative interface. Swap providers by changing two lines of YAML.

Haiku for low-latency replies

Voice agents pay for latency in user perception. Running on Haiku keeps the time-to-first-token short, which matters more for spoken conversations than for written ones.

system prompt for spoken cadence

Agents tuned for written output produce too-long answers in voice. The prompt explicitly asks for sub-30-second responses and conversational pacing.

Read the deep dive

10 apps you can ship in 50 lines of YAML

Read article

Newsletter

Get the next post in your inbox.

Engineering notes from the Digitorn team. No marketing, no launch announcements, no "10 prompts that will change your life". Just the things we write that we'd want to read.

Other templates

developerStarter coder/templates/starter-coder developerClaude Code clone/templates/claude-code-clone researchDeep research agent/templates/deep-research messagingSlack helper/templates/slack-helper developerPull request reviewer/templates/pr-reviewer opsWeekly digest reporter/templates/weekly-digest knowledgeDocumentation generator/templates/doc-writer creativeLive React app builder/templates/react-builder opsScheduled data pipeline/templates/data-pipeline