Choose a provider

Choose OpenAI, ElevenLabs, or Deepgram, then install its package beside @voiceinput/react. Your provider handles transcription and billing; VoiceInput handles field interaction.

Provider	Package	Server key	Setup
OpenAI	@voiceinput/openai	OPENAI_API_KEY	OpenAI guide
ElevenLabs	@voiceinput/elevenlabs	ELEVENLABS_API_KEY	ElevenLabs guide
Deepgram	@voiceinput/deepgram	DEEPGRAM_API_KEY	Deepgram guide

All integrations use an authenticated server route to exchange the long-lived key for a temporary or single-use browser credential. The browser then streams audio directly to the provider. See how it works.

Provider differences

Behavior	OpenAI	ElevenLabs	Deepgram
Default model	gpt-transcribe	scribe_v2_realtime	nova-3
PCM16 sample rate	24 kHz	16 kHz	16 kHz
Omitted language	Automatic	Automatic	multi on multilingual Nova models; otherwise required
Vocabulary	Prompt, or keywords for live models	Key terms	Nova-3 key terms
Omitted endpointing	Server VAD, 500 ms silence	VAD, 650 ms silence	Provider default
endpointing: false	Manual commit	Manual commit	Disables endpointing

gpt-live-transcribe requires endpointing: false. The OpenAI default supports server phrase boundaries.

Adapters validate shared options before asking for microphone permission. Malformed settings produce invalid-configuration; settings the selected provider cannot honor produce unsupported-feature.

Switch providers

Change the package, browser factory, server handler, and server environment variable. The hook, field state, and button stay the same. When overriding a model, configure the same model in the browser factory and server handler.

Provider language support, pricing, latency, and retention differ. Read the selected package guide before overriding language, vocabulary, endpointing, or model settings. To integrate another service, write a custom provider.