Build a Voice Note Transcriber

By the end of this guide you'll have an endpoint your users can send a voice note to and get back a clean transcript — with speaker labels, and in African languages if you need them. Perfect for WhatsApp voice notes, voicemails, or meeting recordings. About ten minutes if you already have a Next.js app.

We'll use the Voice API's transcribe and diarize methods. The provider is inferred from the model, so switching between Deepgram (fast English), AssemblyAI (long-form + speakers), and Spitch (African languages) is a one-line change.

{
  "text": "Hey, just following up on the invoice. Can you send it over today?",
  "duration": 4.2,
  "language": "en",
  "provider": "deepgram",
  "segments": [
    { "speaker": "speaker_0", "text": "Hey, just following up on the invoice.", "start": 0.1, "end": 2.0 },
    { "speaker": "speaker_1", "text": "Can you send it over today?", "start": 2.3, "end": 4.2 }
  ]
}

• A Cencori project + API key (starts with csk_...)
• npm install cencori

Create an API route that accepts an uploaded audio file and returns the transcript.

// app/api/transcribe/route.ts
import { Cencori } from 'cencori';
 
const cencori = new Cencori({ apiKey: process.env.CENCORI_API_KEY! });
 
export async function POST(req: Request) {
  const form = await req.formData();
  const file = form.get('file') as File;
  const language = (form.get('language') as string) || 'en';
 
  const result = await cencori.voice.diarize({
    audio: await file.arrayBuffer(),
    filename: file.name,
    model: 'assemblyai-universal', // long-form + speaker labels
    language,
  });
 
  return Response.json(result);
}

That's the whole backend. diarize turns on speaker labels and returns verbose_json with segments.

The provider is inferred from model, so tuning for cost, speed, or language is a one-line swap:

For a Nigerian WhatsApp bot handling voice notes in mixed languages, detect or ask the language and route accordingly:

const model = ['yo', 'ha', 'ig'].includes(language) ? 'spitch-stt' : 'nova-3';
 
const result = await cencori.voice.transcribe({
  audio: await file.arrayBuffer(),
  filename: file.name,
  model,
  language,
});

If you want a mic recorder in your UI without writing one, use the React component — it records, uploads, and renders the transcript (with speaker labels) for you:

import { VoiceRecorder } from 'cencori/react';
 
export function NoteTaker() {
  return (
    <VoiceRecorder
      endpoint="/api/transcribe"
      model="assemblyai-universal"
      diarize
      onTranscript={(text) => console.log(text)}
    />
  );
}

Close the loop — turn your text reply back into audio in the same language:

const { audio } = await cencori.voice.speak({
  input: 'Invoice sent — check your inbox.',
  model: language === 'yo' ? 'spitch-tts' : 'aura-asteria-en',
  language,
});
// send `audio` back as a WhatsApp voice note

Because it runs through the gateway, every transcription is:

• Billed and tracked — per-minute cost logged to your dashboard with the provider used
• PII-redacted — the input guard runs on the request
• BYOK-ready — plug in your own Deepgram/AssemblyAI/Spitch keys per project
• Portable — swap providers by changing the model, no code rewrite

That's a production voice-note pipeline in a single file.