Voice AI
A voice assistant that a person can talk to and be answered in the same breath. We have built exactly this on React Native, including the proxy plumbing the platform quietly requires, without rewriting the app around it.
Is this the right tool
Real-time voice is the right call when hands and eyes are busy or speaking is genuinely faster than typing. It is the wrong call when a tap would have done the job.
Skip voice if a well-placed button already does the job. The strongest voice features are the ones where typing was never realistic in the first place. If you are adding voice because it demos well, that is not a reason your users will feel.
How we build it
Voice AI on mobile is less about the model and more about the audio path and the transport. That is where the real engineering lives, and where most first attempts stall.
On Epiphra we used the Gemini Live API with the native-audio model, streaming PCM audio at 16kHz in and 24kHz out with selectable HD voices. Bidirectional native-audio models answer as the user speaks, which is what makes the exchange feel like talking rather than sending a message and waiting.
This is the part most teams hit and did not expect. React Native's WebSocket implementation is not compatible with the Live API, so a thin Node and Socket.IO proxy bridges the app to the model's WebSocket. We built that proxy on Epiphra; without it, real-time voice on RN simply does not connect.
Live input and output transcription lets the user read what was heard and what was said, which matters for trust and accessibility. We pair the model's own transcription with on-device speech recognition (expo-speech-recognition) so the caption path is resilient.
A voice UI has to know when the user started talking, when to stop its own audio, and how to recover from a dropped connection mid-sentence. Getting barge-in and reconnection right is what separates a demo from something a person will use twice.
Trade-offs and failure modes we design around
What it costs to run
Voice AI has a different cost shape than text. Real-time audio models bill for the duration of the session and the audio in both directions, so the meter runs the whole time a conversation is open, not once per message. A feature where people hold long conversations costs more than one built around short, purposeful exchanges.
On top of the model you are running the proxy: a socket server that stays up, scales with concurrent conversations, and carries its own hosting cost. That is a fixed line that grows with simultaneous users, and it is easy to forget when you only price the model.
We size this against your expected concurrent sessions and average conversation length before building, so the number is a decision you made rather than a bill you discover. The same cost-model formula we use for token spend applies to voice minutes.
Proof
We are precise here: our production voice-AI example is one app, and we will not pad the list with things that are not really voice AI.
Inseed client, led by our founder
A live AI reflection companion you can talk to. We built the real-time voice assistant on the Gemini Live API: bidirectional native-audio conversation with selectable HD voices, live input and output transcription, and the Node and Socket.IO proxy that bridges React Native to the Live API's WebSocket. Our founder led the architecture; the Inseed team implemented the production app.
Built by our founder
Not AI voice, but real-time voice at production scale, which is the harder half of the plumbing. A HIPAA telehealth app where our founder integrated in-app voice and video calling (GetStream) and a native masked-calling bridge (Doximity Dialer) so patients could call doctors without exposing a real number. Live for over two years.
Straight answer on the gap: Epiphra is our one production real-time AI-voice feature. VocalMD proves we can ship real-time voice infrastructure on React Native, but its calling is human-to-human, not a voice model. If your product needs voice AI, we would prototype it against your use case in the audit before you commit to a build, rather than pointing at a pile of demos.
FAQ
Because React Native's built-in WebSocket does not speak the protocol the real-time voice APIs expect. On Epiphra we bridged the app to the Gemini Live API with a thin Node and Socket.IO proxy. It is a small server, but it is not optional: without it the app cannot hold a live audio connection to the model. Any real-time voice feature on RN carries this piece of infrastructure.
Native-audio models that stream both directions can feel conversational, answering while the user is still finishing a sentence. The delay you feel is the sum of the network, the proxy hop, and the model, so we budget latency end to end and test on real devices and real networks, not just a fast office connection. If latency cannot be kept conversational for your users, we will tell you before building.
The interface can be localized broadly; the spoken conversation is a separate question. On Epiphra the UI shipped in six languages while the live voice ran English-only. We are explicit about which languages work out loud for a given model rather than implying the voice speaks everything the UI does.
Yes. The voice assistant is a feature and a proxy that sit beside your app, not a rewrite of it. Your existing screens, auth, and data stay where they are, and the voice layer talks to the model through the proxy. That is the same "no rewrite" approach we take to every AI integration.
Where this goes next
These are the app categories where talking beats typing, plus our pricing so you can size the work.
30 minutes, free. We will tell you honestly whether voice earns its place for your users, what the latency and cost look like, and what we would prototype first. Shuhel will be on the call.
Book a 30-minute call