Examples

Runnable example apps covering every provider, pattern, and use case.

Every example is a standalone Vite app you can run locally. Clone the repository, install dependencies, and start any example:

git clone https://github.com/composite-voice/composite-voice.git
cd composite-voice
pnpm install
pnpm --filter 01-first-voice-pipeline dev

Each example runs on its own port so you can run several simultaneously. Found 49 examples.

Foundation

first-voice-pipeline

secure-api-keys

Secure API keys with Express proxy — SpeechmaticsSTT + AnthropicLLM + SpeechifyTTS via server-side key injection

cloud-providers

Production-quality pipeline: Speechmatics STT + Speechify TTS + Anthropic Claude via Vite proxy

base-provider-options

conversation-history

eager-pipeline

tool-use

custom-logging

Server & Deployment

native-stt

NativeSTT configuration explorer — Web Speech API + Anthropic Claude + NativeTTS

deepgram-stt

DeepgramSTT configuration explorer — Deepgram Nova + Anthropic Claude + NativeTTS

assemblyai-stt

AssemblyAISTT configuration explorer — AssemblyAI + Anthropic Claude + NativeTTS

elevenlabs-stt

ElevenLabsSTT configuration explorer — ElevenLabs Scribe + Anthropic Claude + NativeTTS

deepgram-flux

DeepgramFlux (V2 STT) demo — currently disabled, shows eager pipeline concept

Deepgram Pipelines

anthropic-llm

Anthropic LLM provider — NativeSTT + AnthropicLLM + NativeTTS with model/temperature/token controls

openai-llm

OpenAI LLM provider — NativeSTT + OpenAILLM + NativeTTS with model/temperature/token controls

groq-llm

Groq LLM provider — NativeSTT + GroqLLM + NativeTTS, ultra-fast inference

gemini-llm

Gemini LLM provider — NativeSTT + GeminiLLM + NativeTTS with model/temperature controls

mistral-llm

Mistral LLM provider — NativeSTT + MistralLLM + NativeTTS with model/temperature controls

webllm

WebLLM in-browser LLM — NativeSTT + WebLLMLLM + NativeTTS, no server needed

openai-compatible

OpenAI-compatible LLM — NativeSTT + OpenAICompatibleLLM + NativeTTS, custom endpoint

Provider Showcases

native-tts

Native TTS provider — NativeSTT + Anthropic Claude + NativeTTS with voice, rate, and pitch controls

deepgram-tts

Deepgram TTS provider — NativeSTT + Anthropic Claude + DeepgramTTS with voice, sampleRate, and format options

openai-tts

OpenAI TTS provider — NativeSTT + Anthropic Claude + OpenAI TTS with model, voice, format, and speed options

elevenlabs-tts

ElevenLabs TTS provider — NativeSTT + Anthropic Claude + ElevenLabs TTS with voiceId, modelId, stability, and similarityBoost

cartesia-tts

Cartesia TTS provider — NativeSTT + Anthropic Claude + Cartesia TTS with voiceId, modelId, emotions, speed, and language

express-proxy

Express proxy with full security config — rateLimit, maxBodySize, authenticate, CORS

nextjs-proxy

Next.js App Router proxy — createNextJsProxy with catch-all route and security config

node-proxy

Plain Node.js HTTP server proxy — createNodeProxy with handleRequest + attachWebSocket

microphone-input

Deep-dive into MicrophoneInput config — all AudioInputConfig options with real-time audio level meter

buffer-input

BufferInput for file/programmatic audio — upload a WAV file and feed it through the pipeline

browser-audio-output

Deep-dive into BrowserAudioOutput config — bufferSize, minBufferDuration, sampleRate, enableSmoothing

null-output

NullOutput for headless/testing scenarios — TTS events fire but no audio plays

event-inspector

Advanced event inspector — real-time timeline, payload display, filtering, wildcard subscriptions, queue events

conversation-strategies

Side-by-side comparison of conversation history strategies — maxTurns=3 vs maxTurns=10

error-recovery

RecoveryOrchestrator demo — configure recovery strategy, simulate errors, track recovery attempts

barge-in

Automatic barge-in demo — interrupt the agent mid-speech with stopSpeaking()

backpressure

Pipeline backpressure demo — adjust maxPendingChunks and observe LLM-to-TTS throttling

audio-config

AudioCapture internals — AudioWorklet vs ScriptProcessor detection, audio chunk stats

custom-provider

Build a custom LLM provider — MockLLM with canned responses, no API keys needed

multi-language

Language switching demo — change SpeechmaticsSTT language at runtime with a selector

guardrails

Guardrails demo — async filters between LLM output and TTS (PII redaction, pronunciation, blocklist, moderation)

Agent Providers

deepgram-agent

Deepgram Voice Agent API — single WebSocket handles STT, LLM, and TTS server-side

twilio-phone-agent

Phone agent on Twilio Media Streams — TwilioMediaStream + SpeechmaticsSTT + AnthropicLLM + DeepgramTTS, one pipeline per call

vonage-phone-agent

Phone agent on the Vonage Voice API WebSocket bridge — VonageAudioSocket + SpeechmaticsSTT + AnthropicLLM + DeepgramTTS, one pipeline per call

discord-voice-bot

Discord voice-channel bot — DiscordVoice duplex provider with Speechmatics STT, Deepgram TTS, and Claude

zoom-meeting-listener

Zoom RTMS meeting listener — live transcripts over a plain node:http webhook, with an end-of-meeting Claude summary

webrtc-loopback

WebRTCInput + WebRTCOutput over a local RTCPeerConnection loopback — join anything WebRTC with zero platform accounts

google-meet-listener

GoogleMeetInput (Developer Preview) — transcribe a live Google Meet conference and take LLM notes, receive-only

teams-meeting-agent

TeamsCall duplex — a voice agent that joins a Microsoft Teams meeting via Azure Communication Services, listens, and speaks replies

Where to start

New to the SDK? Start with 00-minimal-voice-agent — browser-native providers, one API key.

Building for production? Jump to 20-deepgram-pipeline for cloud-grade STT/TTS, then add 10-proxy-server for secure key management.

Simplest setup? Try 70-deepgram-agent — one provider handles STT, LLM, and TTS over a single WebSocket.

Privacy-sensitive? See 50-webllm-pipeline — everything runs in the browser, no data leaves the device.