OpenAI.fm Explored: Turning Plain Text Into Lifelike AI Speech

A full afternoon went into clicking through every voice, rewriting the same script a dozen ways, and listening back to each take on OpenAI.fm. What stuck was how little friction sits between a blank text box and a voice that actually sounds like it means what it says. No account. No install. One paste, one tweak to the tone, one press of Play, and the words come back spoken.

This walkthrough breaks down what the tool is, the three controls that matter, the exact flow for using it, and the real output from the test runs. Pricing, a feature scorecard, and a straight verdict sit at the end.

At a glance

What OpenAI.fm Actually Is

OpenAI.fm is a free playground that OpenAI launched on March 20, 2025 alongside its next generation of audio models. It exists for one reason: to let anyone hear the gpt-4o-mini-tts text-to-speech model before wiring it into a real product. Type a script, pick a voice, describe how it should be delivered, and the page speaks it back in a few seconds.

The headline trait of this model is steerability. Older text-to-speech engines let you choose a voice and little else. Here the model reads free-form instructions about tone, pace, accent and emotion, so the same sentence can land as a calm bedtime whisper or a loud carnival barker depending on what gets typed into the instructions.

The Three Controls That Run Everything

The whole screen comes down to three decisions: who speaks, how they speak, and what they say. Getting comfortable with these three is the entire learning curve.

1. Voice

The top row holds a set of named voice cards. Clicking one marks it with a small orange dot, and a shuffle button at the end of the row throws a random pick if decision fatigue sets in. The two newest voices, Cedar and Marin, are the ones OpenAI flags as its best quality for natural reads.

Voices in the demo

The API also offers a few extra voices beyond the demo row, including echo and shimmer, so the production set is slightly larger than what the playground shows.

Quick voice picker by project

2. Vibe

Below the voices sit the vibe presets, ready-made delivery styles plus a refresh button for a fresh random preset. Each preset loads a block of instruction text that can be edited line by line. This is where the model gets told not just what to say, but how to feel while saying it.

Vibe presets

3. Script

The script box on the right is the text that gets spoken. It accepts a healthy chunk of text per run, and small stage directions written inside the script, such as asking a line to be whispered, get treated as further guidance to the model. Short, punctuated sentences read far more naturally than long run-ons.

How It Works, Start to Finish

The Hands-On Test

To push past the stock sample, a fresh script was written separately and dropped into the tool. The brief was a short cinematic voiceover about a cosmic Incursion, the kind of dramatic narration where delivery matters as much as the words. With the script in place, Cedar was selected for its grounded tone and the Old-Timey vibe was loaded to see how far the instructions could bend the read. The instruction block carried detailed guidance on emotion, inflection and word choice, which is exactly the kind of steering the model is built to follow.

Generation was quick, with audio ready in a handful of seconds. During playback the button switches to Stop, so cutting a take short is one click. Cedar delivered the galaxies-colliding passage with weight and natural pauses, and the instructions clearly shaped the pacing rather than being ignored. On an invented proper noun the read wobbled slightly, which matches how the tool handles unusual words and names. Everyday sentences came back clean.

To listen to the full audio generated in this test click here

Test run at a glance

The Developer Side: Output You Can Ship

The small code toggle at the top right is the feature that separates this from a toy. Flip it on and the exact voice, script and instructions turn into a working snippet, with tabs for Python, JavaScript and cURL. Nothing has to be rewritten by hand, because the same settings that produced the test audio carry straight into code.

What that view hands over is a complete, paste-ready call that reproduces the audio heard in the demo. The model name, the Cedar voice, the script and the instruction block are all filled in, so the jump from playground to project is close to zero. The bottom bar swaps in developer actions: Docs opens the official reference, Copy grabs the whole snippet, and Play still runs the audio so the output can be checked against the code.

Audio actions vs code actions

Pricing

The demo itself costs nothing to use. The model behind it, gpt-4o-mini-tts, is billed through the OpenAI API at roughly $0.015 per minute of generated audio. That rate is what makes it attractive for production work rather than just experiments, since a large batch of narration stays inexpensive compared with many commercial voice services.

Pros and Cons

Who Gets The Most Out Of It

Final Verdict