10 AI tools to help your radio station grow

Most “AI for radio” articles describe a category and leave you to go shopping. This one names ten actual products, what each costs as of September 2026, and the specific radio job it does. The useful part is the honesty at the end: eight of the ten have free tiers a one-person station never outgrows, and the whole realistic bill for a serious independent broadcaster is about $20 a month, or about $42 if you make your own imaging.

The organising idea is that AI is not one purchase. It’s a set of small, boring jobs — tagging files, cleaning up a bad mic, cutting a clip, drafting a newsletter — that each used to eat twenty minutes. Buy the tool that does the job you actually have. If you want the strategic version of this argument first, read what AI can and can’t do for a radio station; this guide is the shopping list.

1. Show prep and scripting: the two hours a week you get back

Claude or ChatGPT. Free tiers; paid plans around $20 a month each, checked September 2026. This is the one line most stations should pay for, and it replaces the largest block of unpaid time in independent radio: sitting down on a Sunday night and working out what you’re going to say.

The mistake is asking for a finished script. What you get back is bland, and bland reads as bland on air. The thing these tools are genuinely excellent at is the pass before the script: turning a raw pile of material into a running order, generating fifteen questions for a guest so you can throw away eleven, or checking that the three stories in your talk break aren’t secretly the same story.

Use it like this: paste your last hour’s playlist and your show’s one-line premise, and ask for ten link ideas of under twenty words each, written the way people talk, no adjectives. The word limit is what makes it usable — unconstrained output is always too long for a link over an intro. Then rewrite every line in your own voice before it goes anywhere near a microphone. There’s a fuller workflow in the AI show prep and social guide.

2. Voice and imaging: where AI earns its money fastest

ElevenLabs. Free tier with a monthly character allowance; paid plans from roughly $5 a month, with the tier most stations settle on nearer $22, checked September 2026. Imaging is the job where synthetic voice is unambiguously good enough: station IDs, sweepers, “you’re listening to” beds, promos for a show that airs on Thursday. These are short, factual, and nobody expects them to be a person.

Use it like this: write twelve liners in one sitting, render them all, and drop them into rotation as a category so your automation sprinkles them through the day. One liner rendered once sounds like a robot. Twelve rendered in the same voice sounds like a station with a voiceover artist on retainer. If your platform supports rule-based rotation, give imaging its own category with a minimum separation so it never stacks up back to back.

A practical constraint worth knowing before you subscribe: cloning a voice that belongs to a real person needs that person’s explicit consent, and the serious vendors enforce it with a verification step. That’s a feature, not an obstacle. The detail on voice selection, pronunciation control and how to make synthetic reads sit properly in a music bed is in AI imaging and voice for radio.

3. Music metadata: the unglamorous fix that improves every hour

MusicBrainz Picard. Free, open source. This is the tool nobody writes listicles about and the one that fixes the most visible problem. If your library came from ten different sources, half your files have the artist in the title field, “Track 04” where a song name should be, or an album tag that says the name of a folder on a laptop you no longer own.

That matters more than it sounds. Bad tags flow straight through to your now-playing display, your public page, your rotation rules and your play reports. A rotation rule that says “don’t repeat an artist within two hours” cannot work if the artist field is blank on a third of the library.

Use it like this: point Picard at a copy of your library — never the only copy — and use acoustic fingerprint lookup rather than filename matching. It listens to the audio and identifies the recording, which is why it fixes files whose names are useless. Then spot-check a hundred rows by hand before you re-upload; automatic matching is very good and occasionally confidently wrong. Once the tags are clean, everything downstream in your music library and playlists gets better at once.

4. Cleanup, mastering and transcription

Adobe Podcast Enhance. Free, checked September 2026. Upload a speech recording made in a kitchen and it comes back sounding like it was made in a booth. It is genuinely startling on phone interviews and remote guests. It is also aggressive: it will happily strip the room out of something you wanted room in, and it does unpleasant things to music, so use it on voice only.

Auphonic. Free tier of roughly two hours of audio a month; paid credit packs beyond that, checked September 2026. Different job, often confused with the one above. Auphonic does loudness normalisation and levelling — it makes your recorded show hit a consistent broadcast loudness target so it doesn’t arrive four decibels quieter than the song before it. If listeners reach for the volume knob when your show starts, this is the fix, and two hours a month covers a weekly show.

Whisper. Free and open; desktop wrappers such as MacWhisper around a $40 one-off, checked September 2026. Transcription that is accurate enough to publish. Every recorded show should be transcribed, because the transcript is the raw material for four other things: show notes, a searchable archive, social captions, and the text a search engine can actually read. If you record shows and publish a podcast, the episode description almost writes itself from the transcript.

Use it like this: transcribe, then hand the transcript to your assistant with the instruction “list the six moments a listener would clip, with timestamps, and one sentence each on why”. That output is simultaneously your show notes and your social plan, from one recording, in about four minutes.

5. Clips, artwork, translation, listener email and analytics

Opus Clip. Free tier with monthly credits; paid from roughly $29 a month, checked September 2026. It watches a long recording and proposes vertical clips with captions burned in. Its judgement about what’s interesting is mediocre; its speed at the mechanical part — reframing, captioning, exporting nine-by-sixteen — is excellent. So let it cut, and you choose. Feed it the timestamps you got from your transcript rather than letting it pick, and the quality problem goes away.

Canva Magic Studio. Free tier; Pro around $15 a month or about $120 a year, checked September 2026. Show covers, social templates, a header for your public station page. The free tier is enough for a long time. The thing worth paying for eventually isn’t the AI generation — it’s the brand kit, which locks your colours and fonts so six months of graphics look like one station instead of six.

DeepL. Free tier with a monthly character allowance; Pro from roughly $9 a month, checked September 2026. If a meaningful slice of your audience speaks another language — and for diaspora stations that’s most of the audience — translating your station description, show notes and listener emails is a small change with a disproportionate effect. DeepL reads more naturally than the general assistants for European languages. Have a native speaker check anything that goes on air.

Mailchimp with its built-in assistant. Free plan capped at a few hundred contacts; paid plans from roughly $13 a month and rising with list size, checked September 2026. A weekly “here’s what’s on this week” email is still one of the most reliable ways to get people to actually tune in, and drafting it is the part that doesn’t get done. Draft with AI, then cut it in half by hand — AI newsletters are always too long. More on the surrounding tactics in how to get radio listeners.

Google Gemini. Free tier; paid tier around $20 a month, checked September 2026. Not because it’s better than the others, but because most stations open their listener analytics, see numbers, and close the tab. Type the week’s figures in and ask a real question — “which hour lost the most listeners, and what changed in that hour?” — and you get a starting point for a programming decision instead of a dashboard you feel vaguely guilty about.

6. Which of the ten are actually worth paying for

Here is the part the vendor pages won’t tell you. Of the ten, exactly two are worth money for a station in its first year, and one of those is conditional:

• Pay: one general assistant, around $20 a month. It touches prep, scripts, sponsor copy, show notes and analytics. It is the only line here that pays for itself in week one.
• Pay if you make imaging: ElevenLabs, around $22 a month. Worth it the moment you’re producing more than a handful of liners a month. Below that, the free tier covers you.
• Free tier is genuinely enough: Picard, Adobe Podcast Enhance, Whisper, Canva, DeepL, Gemini, and Auphonic if you produce one show a week. Seven of the ten. That isn’t us being generous — it’s that these products are competing hard and the free tiers are unusually good right now.
• Don’t buy yet: Opus Clip paid, Mailchimp paid. Both are priced for volume you don’t have on day one. Buy Opus Clip when you’re posting clips weekly and the free credits run out; buy Mailchimp when your list is big enough that the free contact cap is the thing stopping you.

So the honest total: about $20 a month for a talk station, about $42 for a music station that makes its own imaging. Everything else on the list is free. That sits alongside your hosting rather than replacing it — Nobex plans are $19 (Starter), $49 (Pro) and $99 (Business) a month with a 7-day money-back guarantee, and every new account starts in free test mode with a real station on air and no card required.

7. Listen before you air it, every single time

Synthetic voice fails in a specific and predictable way: it is fluent right up until it isn’t, and the failure lands on exactly the words your audience cares most about. Local place names. Your station’s own name. Artist names. Times and phone numbers. A model that reads three sentences beautifully will then say your town’s name with the stress on the wrong syllable, and every local listener will notice, because that’s the word they know best.

There is no automated check for this. The only reliable one is a person with headphones, and it takes seconds:

1. 1Play the rendered file end to end, out loud, at normal volume. Never approve a file by looking at the waveform.
2. 2Check the proper nouns specifically — your station name, presenter names, place names, artist names, sponsor names. These are where models fail and where listeners are least forgiving.
3. 3Check every number: times, dates, frequencies, phone numbers, prices. “1800” read as a year instead of a time makes a promo useless.
4. 4Listen for the register. A model reading a cancellation, a memorial or a closure in a bright commercial-voice tone is the single worst thing AI does on air.
5. 5Play it once against a music bed at real levels. Synthetic voices often sit differently in a mix than a human read of the same script, and something that was clear in isolation can vanish under a bed.
6. 6Only then add it to rotation. Then listen to the station for an hour and hear it in context, where a liner that was fine alone can turn out to be jarring after a ballad.

8. Where AI must not touch your station

Every tool above is a production shortcut. None of them should touch the three things below, and the reasoning isn’t squeamishness — it’s that each one puts something at risk that you can’t buy back.

Anything that misrepresents a human host

Cloning a presenter’s voice to cover a shift they didn’t work, or generating a host who doesn’t exist and letting listeners believe otherwise, breaks the one thing radio has that other media don’t: the audience’s belief that someone is actually there. Synthetic imaging in a neutral station voice is fine and universally understood. A synthetic *person* presented as real is a different product, and when it comes out — it always comes out — the damage is to the trust, not the recording. If a voice on your station isn’t a person, don’t build the audience relationship as though it is.

AI-generated “artists” passed off as real recordings

Generated tracks credited to invented artists and slotted into rotation as though they were signed acts is a practice several platforms have had public trouble over, and it damages you twice. Listeners who discover it stop trusting your music selection. And your reporting becomes fiction: the play logs behind your metadata and royalty reporting are supposed to describe real recordings by real rightsholders, and filling them with invented credits corrupts the one record that proves what your station actually played.

Anything that touches your rights position on music

This is the one to be most careful about. The copyright status of AI-generated music is unsettled and differs by country, and “the tool said I own it” is not a rights position. If you use generated music as beds, imaging or filler, keep it in a clearly separated category in your library with the source recorded, so that the boundary between “music we hold rights to play” and “audio we generated” is documented rather than remembered.

More broadly: a licence bundled by a hosting provider sits in the host’s name and is scoped to their territories, so it doesn’t travel with you when you leave. A licence in your own name is an asset of your station. Hosting and licensing are separate purchases and should stay that way. The detail is in the internet radio licensing guide and how to get a music license. Nobex keeps automated playout inside the statutory playout rules — the cloud AutoDJ enforces the US sound-recording performance complement for you — but broadcasters bring their own licences, and no AI tool changes that.

9. Questions people actually ask