Narration and audio content production — Grok Voice & Audio

Prices as of November 3, 2026, from the xAI Pricing page: TTS costs $15 per million characters. Source: Pricing.

Turn text into professional audio content

Until recently, producing audio at scale meant a studio: a voice actor, a booth, a sound engineer, and a price tag that limited voiced content to your flagship pieces. Grok TTS moves that line. Automated podcasts, articles read aloud, training modules, voice notifications: the quality-to-cost ratio makes viable the kind of projects people used to drop as soon as they saw the quote.

Podcasts and audio newsletters

An automated podcast pipeline comes down to four steps:

Writing: your LLM (Grok or another) writes the podcast script, including expressive tags
Production: TTS generates the audio with the voice and effects you chose
Post-production: adding an intro/outro jingle, normalizing the volume
Publishing: uploading to your distribution platform

The quality of the result is decided almost entirely in step 1. A script written like an article sounds like someone reading an article; a script written for the ear, with breaths and changes of pace, sounds like a show. Here’s what the start of an annotated episode looks like:

[breath] Hi, everyone, and welcome to your weekly tech roundup. [pause]

We've got three topics on the agenda today. [long-pause]

<emphasis>First</emphasis>, the latest Grok API update. [pause]
<emphasis>Second</emphasis>, the new TTS features. [pause]
And <emphasis>third</emphasis>, lessons learned from running the Voice Agent in production. [long-pause]

<build-intensity>Let's jump right into the first topic.</build-intensity>

Your choice of voice then sets the show’s register. Rex suits a professional tech podcast, Eve a dynamic, high-energy format, Ara an intimate, conversational tone, and Leo an educational show where listeners expect calm authority. Among the newer voices, xAI lists orion and lux for narration and cosmo for podcasts. Listen to three or four candidates reading the same paragraph of your script before you decide: with thirty seconds of real text, the right choice is usually obvious.

E-learning and training

Every lesson in a course can come with an audio version, so learners can follow it “podcast style” on their commute or on a walk. The economics are compelling: a 10,000-word course, roughly 60,000 characters, costs about $0.90 to voice. At that price, the question is no longer which modules deserve an audio version, but whether there’s any reason not to voice them all.

The next step is to combine the two building blocks. TTS reads the exercise instructions, the Voice Agent listens to the learner’s answer, and the model evaluates it and gives immediate spoken feedback. You get an oral practice partner available around the clock, especially useful for languages, sales pitches, or interview prep.

Then there’s accessibility, which is often a legal requirement rather than a bonus. TTS makes your written content usable by people with low vision or dyslexia, and a “Listen to this article” button on your pages generates the audio on the fly, with no prior production or storage.

Voice notifications and alerts

Some messages have to reach people who aren’t looking at a screen. A server that goes down at 3 AM triggers a voice alert on the on-call engineer’s phone; an assistant reads out the day’s appointments; an automated call confirms an order by recapping its details.

These messages are short, so HTTP mode is more than enough:

import requests

def generate_voice_alert(message, language="en"):
    response = requests.post(
        "https://api.x.ai/v1/tts",
        headers={"Authorization": f"Bearer {API_KEY}"},
        json={
            "text": message,
            "voice_id": "leo",
            "language": language,
            "output_format": {"codec": "wav", "sample_rate": 8000}
        }
    )
    return response.content  # raw audio bytes

WAV at 8000 Hz isn’t a quality choice but a compatibility one: it’s the format phone systems expect, and delivering it directly avoids a transcoding step that would degrade the signal.

Audiobooks and long-form narration

For a book or a report dozens of pages long, switch to WebSocket mode. The method: split the text into chapters or sections, send each section with text.delta, collect and concatenate the audio chunks you receive, then add metadata to the final file, such as chapter markers and a table of contents.

Voice consistency then becomes an issue in its own right, because listeners spend hours with the narrator and notice the slightest break. Keep exactly the same voice for the same narrator from start to finish. Insert [long-pause] between chapters to mark the transitions the eye used to get from the layout. And vary the pace where the meaning calls for it: <slow> on a dense passage that needs time to sink in, <fast> on a transition or a secondary list.

Key takeaways

TTS makes large-scale audio production possible: podcasts, e-learning, notifications, audiobooks
Expressive tags bring narration to life with pauses, changes of pace, and vocal effects
At $15 per million characters (xAI pricing as of November 3, 2026), the cost is modest compared with studio recording
Combine TTS (narration) and the Voice Agent (interaction) for complete learning experiences
For long texts, use WebSocket and split the text into sections