Raw WebSocket

The framework-agnostic recipe. If your agent, telephony provider, or pipeline can hand you raw audio frames, you can add a Voxtar diagnostic layer — no SDK required. Every named framework recipe is a thin wrapper over exactly this.

The endpoint is /analyze/stream. Open it once per call, send a start frame, forward audio, and read events.

Audio requirements

Send pcm_s16le (16-bit signed little-endian PCM), mono, 16kHz or higher — the same format most telephony media streams already emit. If your source is Opus or µ-law, decode/expand to PCM first, or set encoding accordingly (opus, mulaw). Frames of 20–100ms are ideal.

Minimal client

import { WebSocket } from "ws";
export function attachVoxtar(patientId, audioSource) {
  const vox = new WebSocket(
    "wss://api.voxtar.hallelx2.com/v1/analyze/stream",
  );
  vox.on("open", () => {
    vox.send(JSON.stringify({
      type: "start",
      api_key: process.env.VOXTAR_API_KEY,
      patient_id: patientId,
      encoding: "pcm_s16le",
      sample_rate: 16000,
      panels: ["respiratory", "neurological"],
      write_to_ehr: true,
    }));
    // Forward raw PCM frames from whatever produces the patient's audio
    audioSource.on("frame", (pcm) => vox.send(pcm));
    audioSource.on("end", () => vox.send(JSON.stringify({ type: "stop" })));
  });
  vox.on("message", (raw) => {
    const evt = JSON.parse(raw.toString());
    if (evt.type === "partial") {
      // live readings every ~1s — optional to use
    }
    if (evt.type === "complete" && evt.flags.length) {
      notifyCareTeam(patientId, evt.flags);
    }
  });
  return vox;
}

Events

You'll receive partial readings every ~1s and one complete event when you send stop (or close the socket), carrying the same shape as POST /analyze. The completed analysis is persisted and retrievable via GET /analyses/{id}.

Adapting to your stack

The only per-framework difference is where you get audioSource: LiveKit gives you an rtc.AudioStream, Twilio a Media Streams WebSocket, Vapi/Retell a call audio hook, Pipecat a frame processor. Wire any of those into the client above and you're done.