I produce a Saturday show for a volunteer-run community station in Greensboro. Our budget for music beds, intros and event promos is roughly zero, so most weeks I recycle the same royalty-free loops everyone in the Triad uses.
So one afternoon this month I tested an AI music model on what we actually need: a show opener, a station ID sting, a bed to talk over, and a promo for a neighborhood porch-music night (invented for the test). I timed every request and measured every file. Total spend: 60 cents.
The setup
The model I tested is called Flow Music. You type a style description, and optionally your own lyrics, a tempo in beats per minute, a target length and a seed, and it returns a finished track. There are two versions listed, Lyria 3 Pro (the default) and Lyria 3.5, and both cost the same at the time of testing: $0.06 per generation request on the pay-as-you-go page, versus a listed official price of $0.075. The same settings appear in a simple web form, so nobody needs to write code to try it.
I made 10 requests across six tests. Every one completed. None failed or had to be retried.
Tests 1 and 2: the show opener, and the length problem
My first prompt asked for an upbeat Piedmont-style string band instrumental, “fingerpicked acoustic guitar, banjo, upright bass, light washboard percussion, warm and friendly, no vocals,” at 112 BPM and 15 seconds long.
It took 43 seconds to come back. The track was 133 seconds long.
That was the most consistent finding of the day. The length setting is documented only as a “target,” and in practice it was ignored every time. Across 11 clips I asked for 15 to 60 seconds and got 61 to 181 seconds, between 2.1 and 8.9 times what I asked for.
Running the same opener prompt on Lyria 3.5 at 30 seconds gave me two clips for the one 6-cent charge (66 and 122 seconds) and was the slowest request of the day at 78 seconds. A follow-up “15-second station ID sting” that said “ends cleanly with a final chord” in the prompt itself still came back at 61 seconds.
The fix is boring and fine: open the file in a free editor and cut on a bar line near the mark you need. The music holds a steady level with no mid-track dropouts (the opener did end with 11 seconds of silence and a stray blip), so there is always material to trim.
Test 3: a bed you can talk over
For our interview segment I asked for “soft Rhodes electric piano, brushed drums, warm bass, relaxed late-night jazz feel, sits under a talking voice, no vocals, no lead melody that competes with speech.” It came back in 60 seconds at 176 seconds long.
Speech intelligibility lives mostly between 1 and 4 kHz. In the bed, only 0.3% of the energy fell in that band; in the sung promo from Test 4, 26.5%. That bed leaves room for a host’s voice.
Test 4: a promo with real lyrics
I wrote four lines for the invented porch-music night: “Friday night on the east side / Bring a chair and bring a friend / Porch Night, Porch Night / Music on the corner till the evening ends.” I asked for Southern soul and funk with a horn section, a female lead and call-and-response backing singers, at 104 BPM, on Lyria 3.5.
All four lines were sung in order and the chorus came twice. The model also added an outro line I did not write (“Yeah, we’re feeling good tonight”) plus backing-vocal echoes. An automatic transcription heard “March night” instead of “Porch night” on the chorus. In a promo, the event name is the one word that has to land, so I’d want a human singer to re-record it. The track ran 62 seconds against the 30 I requested.
For a station that just needs a rough jingle demo to pitch to a sponsor or a venue, the Flow Music API got me there in 43 seconds for 6 cents. That is a draft, not a finished spot.
Test 5: genre spread
Three styles that matter on our airwaves: Piedmont blues (solo fingerpicked guitar with harmonica fills), a small-club jazz trio, and a boom-bap hip-hop beat. Each was asked for 30 seconds without vocals. They came back at 147, 177 and 181 seconds, in 43 to 52 seconds of waiting.
Tempo was the most reliable control. With a rough beat-tracking script I measured every clip, and they all landed within about 4 BPM of what I asked for. One flag: the jazz trio’s metadata listed a hummed intro (“Hmm, hmm, hmm”) even though I wrote “no vocals.” Check every “instrumental” before you put it under someone’s voice.
Test 6: can you get the same track twice?
The documentation says the seed setting gives “similar results” but exact repeats are not guaranteed. I ran the same Americana bed prompt twice with seed 4242. Both came back at 129.7 seconds, and the two audio files matched at a 0.99999 correlation, effectively identical. Handy if a file goes missing, though results are only stored for 72 hours. Loudness was consistent too: every clip measured -14 to -16 LUFS integrated.
What it costs, honestly
Ten requests, 11 clips, 60 cents. Average wait 49 seconds. By the numbers (tempo on target, steady level) all 11 are workable as drafts, and the promo works as a demo. None were broadcast-ready as delivered, because every single one had to be cut to length.
Don’t let this replace the people who play here
The Triad has a deep bench of working musicians, from Piedmont blues players to jazz students, gospel choirs and home-studio hip-hop producers. A 6-cent track should not take a paid gig from any of them.
So I’ll treat AI tracks as temp tracks: placeholders while a volunteer builds a segment, or demos that show a local artist the mood we want before we pay them for the real thing. For anything that runs regularly, like our main show theme, we will commission someone local and credit them on air.
Two more rules. Check the terms of service for commercial and broadcast use before anything airs, because those terms, not a blog post, decide what you are allowed to do. And tell listeners. A short “music bed generated with AI” line in the show notes costs nothing and keeps the trust a community station runs on.
The real value is speed: I can audition ten moods in the time it takes to dig through a loop library. The final cut should still sound like the people who live here.
Join the First Amendment Society, a membership that goes directly to funding TCB‘s newsroom.
We believe that reporting can save the world.
The TCB First Amendment Society recognizes the vital role of a free, unfettered press with a bundling of local experiences designed to build community, and unique engagements with our newsroom that will help you understand, and shape, local journalism’s critical role in uplifting the people in our cities.
All revenue goes directly into the newsroom as reporters’ salaries and freelance commissions.
Leave a Reply