Quickstart (HTTP)

HTTP is great for simple, cost‑efficient transcription. The live aggregator batches PCM frames into WAV and flushes by timer or at the end of a speech segment.

What you’ll build

• Start the recorder and gate audio by speech
• Send HTTP chunks by timer or on segment end
• Receive final transcripts

Tip For “one request per phrase”, set intervalMs = 0 and flushOnSegmentEnd = true.

1) Install

# Required
pnpm add @saraudio/runtime-browser @saraudio/deepgram

# Optional stages (VAD + Meter)
pnpm add @saraudio/vad-energy @saraudio/meter

2) Provider (Deepgram example)

Use a raw API key for the quickest setup (ideal for server or local tests).

import { deepgram } from '@saraudio/deepgram';

export const provider = deepgram({
  model: 'nova-3',
  auth: {
    apiKey: '<DEEPGRAM_API_KEY>',
  },
});

Note: For production browsers, prefer issuing short‑lived tokens from your backend. We’ll cover this in the Auth guide.

3) Controller in HTTP mode

import { createRecorder, createTranscription } from '@saraudio/runtime-browser';
import { vadEnergy } from '@saraudio/vad-energy';
import { meter } from '@saraudio/meter';

const recorder = createRecorder({
  stages: [vadEnergy({ thresholdDb: -50 }), meter()],
  segmenter: true,
});

const ctrl = createTranscription({
  provider,
  recorder,
  transport: 'http',
  flushOnSegmentEnd: true, // enable segment‑only semantics
  connection: {
    http: {
      chunking: {
        intervalMs: 0, // no timer → flush on segment end (or forceEndpoint)
        minDurationMs: 700, // ignore tiny bursts on timers (not used when interval=0)
        overlapMs: 500, // tail continuity between chunks
        maxInFlight: 1, // limit concurrent HTTP requests
        timeoutMs: 10000, // per‑flush timeout
      },
    },
  },
});

4) Start and flush

ctrl.onTranscript((r) => console.log('final:', r.text));
ctrl.onError((e) => console.error(e));

await recorder.start();
await ctrl.connect();

// Optional: manual flush at any time (e.g., on button click)
// await ctrl.forceEndpoint();

How VAD gating works

• With flushOnSegmentEnd: true, the controller subscribes to speech frames only → silence is dropped.
• If you disable VAD or don’t end a segment, call forceEndpoint() to flush in‑flight audio.

Next

• Concepts → Controller & Transport (HTTP chunking in detail)
• Providers → Deepgram / Soniox (HTTP support and models)