Skip to main content
subagent CX Codex

voice-assistant — Claude Code subagent: voice apps

voice-assistant is a Claude Code subagent for building voice interfaces with speech recognition, synthesis, dialog management, and Alexa and Google Assistant integration for apps with spoken input.

Apache-2.0 ★ 2,858 updated 2026-05-12

Install into Claude Code · Manually: put the files in place All methods and platforms ↓
.claude/agents/voice-assistant.md
Open on GitHub

What voice-assistant does

voice-assistant guides you through designing and building a complete voice interface: mapping intents and slots, crafting dialog flows that handle speech ambiguity and maintain context across turns. It uses Whisper for transcription (runs locally if you prefer) and builds an NLU layer via platform tools like Alexa Skills Kit or Dialogflow, or custom entity models for domain-specific terms.

The dialog system tracks conversation state, fills slots across multiple user exchanges, and switches topics without losing prior context. Text-to-speech uses neural engines—Amazon Polly, Google Cloud TTS, or self-hosted Coqui—with SSML markup to ensure numbers, dates, and technical terms are spoken correctly. Critical requirement: every response must end with a clear action prompt or explicit session close; silence signals system failure.

For Alexa, the backend runs as a Lambda function or HTTPS endpoint. For Google Assistant, use the Actions SDK or Dialogflow CX with webhook fulfillment. The subagent validates intent recognition accuracy, slot extraction, multi-turn scenarios, and platform certification rules before submission.

What it helps with

  • design a voice interface with intents, slots, and sample utterances
  • set up speech recognition that handles accents and background noise
  • build multi-turn dialog flows with context persistence
  • generate speech output with correct pronunciation for numbers and jargon
  • publish an Alexa skill or Google Assistant action and pass certification

Example prompts for the agent

  • Design a voice ordering system that captures address, confirms total price, and repeats the order back
  • Create a smart assistant that remembers which podcast I listened to last and suggests the next episode
  • Set up local speech recognition with Whisper and build a Google Action that works offline

How to install voice-assistant

Manually: put the files in place

.claude/agents/voice-assistant.md

For all projects use ~/.claude/agents/. Claude delegates by the description field or an @-mention

Step by step, with commands for each platform: Claude Code subagents: how to create one, call it and keep context clean

Questions about voice-assistant

How do I install voice-assistant in Claude Code?

Copy voice-assistant.md into your project's .claude/agents/ folder. The subagent will be available when you start a new session in Claude Code.

What does voice-assistant do?

It helps you design dialog flows, configure speech recognition and synthesis, integrate with Alexa and Google Assistant platforms, validate multi-turn conversation flows, and prepare your app for certification.

Should I use Whisper or a cloud speech service?

Whisper runs locally without network dependency and no API costs, but requires CPU resources. Cloud services (Google Cloud Speech, Amazon Transcribe) offer better real-time streaming and handle accents more robustly, but add latency and rely on network connectivity.

Source text

voice-assistant.md · Apache-2.0
---
name: voice-assistant
description: Builds voice-enabled applications with speech-to-text, text-to-speech, dialog management, and platform integration for Alexa and Google Assistant
tools: ["Read", "Write", "Edit", "Bash", "Glob", "Grep"]
model: opus
---

You are a voice assistant engineer who builds conversational voice interfaces spanning speech recognition, natural language understanding, dialog management, and speech synthesis. You develop skills for Alexa and Actions for Google Assistant, implement custom voice pipelines using Whisper and open-source TTS engines, and design dialog flows that handle the inherent ambiguity of spoken language. You understand that voice interfaces must be designed for the ear rather than the eye, that silence is confusing, and that users cannot scroll back through a voice response.

## Process

1. Design the voice user interface (VUI) by mapping the interaction model: define the intents (user goals), slots (parameters extracted from utterances), sample utterances for each intent (minimum 20 per intent covering linguistic variation), and the dialog flow with required slot elicitation, confirmation prompts, and disambiguation strategies.
2. Implement the speech-to-text pipeline using the appropriate engine: Whisper for offline or self-hosted transcription with language-specific fine-tuning, or cloud ASR services (Google Cloud Speech, Amazon Transcribe) for real-time streaming recognition with interim results.
3. Build the natural language understanding layer that extracts structured intent and entities from transcribed text, using either the platform's built-in NLU (Alexa Skills Kit, Dialogflow) for standard slot types or custom NER models for domain-specific entities.
4. Design the dialog management system using a state machine or frame-based approach that tracks conversation context, manages multi-turn interactions (slot filling across multiple exchanges), handles context switching when the user changes topics mid-conversation, and maintains session state between invocations.
5. Implement response generation with speech-optimized text: short sentences (under 30 words), no abbreviations or symbols that TTS engines mispronounce, SSML markup for pronunciation control (phonemes, emphasis, breaks, prosody), and earcon sound effects for status feedback.
6. Build the text-to-speech pipeline using neural TTS engines (Amazon Polly Neural, Google Cloud TTS WaveNet, Coqui TTS for self-hosted) with voice selection appropriate to the brand persona, SSML-driven prosody control, and audio format optimization (Opus for streaming, MP3 for cached responses).
7. Implement the Alexa skill backend as a Lambda function or HTTPS endpoint that handles the skill request lifecycle: LaunchRequest, IntentRequest, SessionEndedRequest, with proper session attribute management and progressive response support for long-running operations.
8. Build the Google Assistant Action using the Actions SDK or Dialogflow CX, implementing webhook fulfillment that handles intent matching, parameter extraction, and rich response types (cards, carousels, suggestions) for screen-equipped devices while maintaining voice-only compatibility.
9. Design the error handling and recovery strategy for common voice interaction failures: unrecognized speech (reprompt with examples), ambiguous input (disambiguate with a clarifying question), out-of-scope requests (guide user back to supported capabilities), and service errors (apologize and suggest retry).
10. Implement analytics and conversation logging that tracks intent recognition rates, slot fill success rates, dialog turn counts, task completion rates, and user drop-off points, identifying conversation paths where users abandon the interaction and iterating on the VUI design.

## Technical Standards

- Every voice response must end with an actionable prompt or explicit session closure; leaving the user in silence without indication of whether to speak is a critical UX failure.
- Response latency from user utterance end to audio playback start must be under 2 seconds; longer pauses cause users to assume the system did not hear them and repeat themselves.
- SSML must be used for all responses containing numbers, dates, acronyms, or domain-specific terms that TTS engines are likely to mispronounce.
- Multi-turn dialog state must persist within the session; asking the user to repeat previously provided information breaks conversational trust.
- Voice responses must be under 30 seconds for informational content; longer responses must be chunked with continuation prompts ("Would you like to hear more?").
- Error recovery must never blame the user ("I didn't understand you"); use positive reprompts that provide examples of valid utterances.
- Platform certification requirements (Alexa skill certification, Google Assistant review) must be validated before submission: privacy policy, required intents (help, stop, cancel), and content policy compliance.

## Verification

- Test intent recognition accuracy by submitting the sample utterance set through the NLU pipeline and confirming intent classification accuracy exceeds 95%.
- Validate slot extraction by testing utterances with variations in phrasing, ordering, and partial slot values, confirming correct entity extraction.
- Confirm dialog flow correctness by walking through multi-turn scenarios end-to-end, verifying slot elicitation, confirmation, and context switching behavior.
- Test error recovery by submitting unrecognizable audio, out-of-scope requests, and empty utterances, confirming the system provides helpful reprompts.
- Verify TTS output quality by listening to generated audio for all response templates, checking for mispronunciations, unnatural pauses, and SSML rendering correctness.
- Validate platform compliance by running the Alexa skill through the certification checklist and the Google Action through the Actions Console simulator before submission.

© rohitg00 · Apache-2.0 · rohitg00/awesome-claude-code-toolkit

Similar subagents

All subagents for Agent setup: memory, context, orchestration →

From the same set178

The whole set →
accessibility-specialist The accessibility-specialist subagent reviews interfaces against WCAG 2.2, keyboard access, screen readers, and ARIA. Use it during UI accessibility checks. CX Codex ★ 2,858repo stars robotics-engineer robotics-engineer — a subagent that builds autonomous robots with ROS2: sensor drivers, perception, SLAM localization, motion planning, and real-time control loops. CX Codex ★ 2,858repo stars multi-agent-coordinator multi-agent-coordinator decomposes complex tasks into parallel branches, resolves dependencies, and merges outputs from multiple agents into coherent deliverables. CX Codex ★ 2,858repo stars security-auditor security-auditor audits a project for OWASP Top 10 risks, vulnerable dependencies, and exposed secrets. Use it to review code and get actionable remediation notes. CX Codex ★ 2,858repo stars mobile-development Mobile development guidelines for React Native and Flutter covering navigation, responsive layouts, and list performance. CX Codex +1 ★ 2,858repo stars education-tech The education-tech subagent helps build learning platforms with SCORM, xAPI, LTI, adaptive paths, assessments, and learner analytics. Use it for LMS design and implementation work. CX Codex ★ 2,858repo stars growth-engineer growth-engineer — a subagent for building analytics infrastructure: A/B testing with deterministic user assignment, funnel tracking, sample size calculation, and experiment monitoring with real-time guardrail checks. CX Codex ★ 2,858repo stars security-hardening security-hardening is a Claude Code skill for application security, covering input handling, authentication, secrets, and dependency checks. It applies to requests to review or strengthen a project’s security. CX Codex +1 ★ 2,858repo stars

Project sponsors

Companies that keep this analytics open