On-Device Voice
for React Native,
99% cheaper than cloud
EdgeSpeech adds listening, speaking, and voice AI to your apps with a pure JavaScript API. All audio processing runs on-device via the Switchboard SDK — no streaming, no sockets, no cloud audio costs.
99% cheaper
EdgeSpeech on-device vs cloud speech-to-speech API at 1,000 conversations/day.
0 lines of native code
Pure JavaScript API — no C++, no native audio modules, no Xcode or Android Studio setup.
3 on-device ML models
SileroVAD for voice detection, Whisper for transcription, and Sherpa for synthesis — all on-device.
Cloud voice AI is ruinously expensive
Traditional voice AI pipelines stream raw audio to the cloud for every conversation. At scale, you're paying to encode, transmit, and transcribe every second of microphone audio — whether the user is speaking or just breathing.
At 1,000 conversations per day (5 minutes each), OpenAI Realtime API costs $3,600/day — that's $1.3M per year just for audio I/O.
- Continuous audio streaming to cloud servers
- Round-trip latency degrades the experience
- Transcript logs create privacy and compliance risk
- Vendor lock-in: tied to one provider's speech model
- WebRTC / socket infrastructure required
EdgeSpeech: all audio stays on the device
EdgeSpeech runs VAD, STT, and TTS entirely on-device using the Switchboard SDK. Your cloud LLM only ever sees text — a few hundred tokens per turn. Audio never leaves the phone.
At the same scale, EdgeSpeech + text-only LLM costs ~$20/day. That's a 180× cost reduction, plus lower latency and zero audio privacy risk.
- On-device VAD, STT, and TTS — no audio streaming
- Works with any LLM (chat pipeline receives text only)
- Pure JavaScript API — no native audio code
- No WebRTC, no sockets, no cloud audio infrastructure
- Works offline — all ML models run locally
The numbers don't lie
1,000 conversations/day · 5 minutes each · at scale
OpenAI Realtime API
Cloud Speech-to-Speech
$3,600/day
- $0.06 / min for audio I/O streaming
- LLM multimodal tokens billed separately
- ~$108,000 / month at scale
- Audio routed through cloud on every call
EdgeSpeech + Text LLM
On-Device Voice AI
~$20/day
- $0 for STT — Whisper runs on-device
- $0 for TTS — Sherpa runs on-device
- Text-only LLM tokens only — ~90% cheaper
- ~$600 / month at scale — 180× savings
Monthly savings at this scale
$107,400/mo
That’s $1.3M saved annually — redirected into product, headcount, or margin.
How it works
EdgeSpeech manages two Switchboard SDK audio graphs. Your JavaScript code sits in the middle — receiving text, calling your LLM, and issuing speak() calls.
Under the hood — Switchboard SDK audio graphs
Microphone → MultiChannelToMono → BusSplitter → SileroVAD + Whisper STT
Sherpa TTS → Speaker Output
All processing: 16-bit, 16kHz audio, on-device. No intermediate API calls for audio.
Built for developer velocity
Everything you need to ship a production voice AI app — without becoming an audio engineer.
On-device processing
VAD, STT, and TTS run entirely on the device via the Switchboard SDK runtime. No audio ever leaves the phone. Zero privacy risk, zero transcript logs.
JavaScript-only API
initialize(), start(), speak() — that's your entire surface area. No native modules to write, no Xcode builds to debug. Just React Native JavaScript.
Works with any LLM
Your chat pipeline receives plain text. Use OpenAI, Anthropic, Gemini, Llama, or a local Ollama model. No vendor lock-in, no special integration required.
No infrastructure needed
No WebRTC, no audio sockets, no cloud audio pipelines. EdgeSpeech is a React Native library — install it like any other npm package.
Ultra-low latency
Local VAD means transcription starts the moment you stop speaking — no round-trip delays. TTS synthesis begins as soon as your LLM returns the first token.
Barge-in support
onInterrupted event fires when the user speaks while the app is talking. Handle interruptions naturally — pause, listen, and respond. Just like a real conversation.
Up and running in 5 minutes
Install the package
npm install @synervoz/edgespeech
Initialize with credentials
Get your App ID and Secret from console.switchboard.audio.
Listen for transcripts
The onTranscript event fires with text. Pass it to your LLM. Your JavaScript receives text in, and returns text out.
Speak the response
Call speak(text) and Sherpa TTS renders audio on-device. No streaming, no buffering — it just plays.
import { initialize, start, speak, SwitchboardVoiceModule } from '@synervoz/edgespeech'; // 1. Initialize with your Switchboard credentials await initialize('YOUR_APP_ID', 'YOUR_APP_SECRET'); // 2. Listen for final transcripts SwitchboardVoiceModule.addListener( 'onTranscript', async ({ text, isFinal }) => { if (!isFinal) return; // skip partials // 3. Send text to your LLM (any provider) const reply = await callYourLLM(text); // 4. Speak the response on-device await speak(reply); } ); // 5. Handle barge-in (user interrupts) SwitchboardVoiceModule.addListener( 'onInterrupted', () => console.log('User interrupted') ); // 6. Start listening await start();
Supported platforms
Built for React Native developers. Native audio engine, JavaScript developer experience.
iOS
Supported
iOS 13.4+ required. Fully supported with example app and full documentation available.
React Native
Supported
React Native 0.74+ and Node.js 20+ required. Pure JS API — no Expo ejection needed.
Android
Coming soon
Android support is in active development. API design is platform-agnostic — your code won't change.
Web / Expo
Planned
Web and Expo support is on the roadmap. Follow the GitHub repository for updates.
Ship your voice feature today
The library is on npm, the source is on GitHub, and you'll have a working example running in under 10 minutes.