Transcribe and narrate

Build an audio-to-text-to-audio pipeline with honest transcript detail, separate Safety boundaries, and explicit persistence.

This recipe transcribes a recording, creates a short narration script, and generates speech. Each stage has its own result and failure boundary.

1. Bind the adapter

import { readFile } from "node:fs/promises";
import OpenAI from "openai";
import { createOpenAI } from "@use-crux/openai";
import { inMemoryAssetStore } from "@use-crux/core/storage";

const openai = createOpenAI(new OpenAI({ apiKey: process.env.OPENAI_API_KEY }));
const assets = inMemoryAssetStore();

2. Transcribe measured detail

const recording = await readFile("./recordings/weekly-sync.wav");

const transcript = await openai.transcribe({
  model: "gpt-4o-mini-transcribe",
  audio: recording,
  language: "en",
  timestamps: "segment",
  timeout: {
    totalMs: 60_000,
    stepMs: 45_000,
  },
});

console.log(transcript.text);
console.log(transcript.segments);

Request timing only when the selected endpoint measures it. Empty detail stays empty; Crux never estimates intervals. If speaker labels are required, select a model that supports diarization and request it explicitly.

3. Derive the narration script

Use an ordinary prompt for language work:

import { prompt } from "@use-crux/core";
import { z } from "zod";

const narrationScript = prompt({
  input: z.object({ transcript: z.string() }),
  prompt: ({ input }) =>
    `Write a concise spoken summary of this transcript:\n\n${input.transcript}`,
});

const script = await openai.generate(narrationScript, {
  model: "gpt-4o-mini",
  input: { transcript: transcript.text },
});

This is intentionally a separate model operation. Its text Safety, constraints, cost, retry, and trace evidence do not get hidden inside transcription.

4. Generate speech

const narration = await openai.generateSpeech({
  model: "gpt-4o-mini-tts",
  text: script.text,
  voice: "alloy",
  instructions: "Clear, neutral, and suitable for a short briefing.",
  outputFormat: "mp3",
});

Speech input and output have their own guardrail boundaries. Apply a policy to the script or audio when the product requires it; do not assume transcription policy automatically governs a later synthesis call.

5. Persist the final audio

const storedNarration = await assets.put(narration.audio);

Persisting the source recording, transcript, and generated narration are three independent application decisions. Crux does not retain one merely because the next stage consumes it.

If storage fails, retry this put() with the existing audio. Do not re-run transcription, script generation, or speech synthesis.

Stream delivery when necessary

Use streamSpeech() when your consumer understands the provider's native audio framing and progressive delivery improves latency. Persist completion.audio, not provisional chunks.

See Speech generation, Transcription, and Storage and delivery.