DEV Community

Paragon
Paragon

Posted on Originally published at paragonaimag.substack.com

How AI Music Generation Works

By Paragon AI MAG — Series 4, Episode 1. Every fact below is checked against public sources; the numbers come from the sources named in the text.

The band that doesn't exist

In June 2025, a band called The Velvet Sundown released two albums within a month and landed on popular Spotify playlists. By July 1 they already had 550,000 monthly listeners. Then they rocketed to a peak near 1.4 million — numbers most independent artists work their whole lives for, built by an algorithm in weeks. No tour, no press, no label budget. No band, either.

On July 5, 2025, the creators admitted it: a "synthetic music project guided by human creative direction, and composed, voiced, and visualized with the support of artificial intelligence." The songs were made with Suno. The audio verification firm Ircam Amplify tested the catalog and flagged 10 of 13 tracks as 100% AI-generated. By October 2026, Spotify had put a "Likely AI Persona" badge on the profile, with about 107,000 monthly listeners left.

So how does a machine turn one typed sentence into a full song — singer, instruments, lyrics and all — and if it is that good, why is almost nobody actually listening to the flood of AI songs?

The short answer: a song is never played

An AI song is never actually played. The model predicts sound the exact same way ChatGPT predicts words — one tiny piece at a time.

The problem is density. Three minutes of music is almost 6 million numbers to a computer (32,000 samples a second in MusicGen). Predicting them one by one is far too slow. So the machine squeezes first. A neural audio codec — EnCodec, from Meta — compresses each second of music into about 200 tokens, roughly a 600x squeeze. The first codebook is a rough sketch; the later codebooks add detail.

Then a transformer model predicts the next audio token, guided by your words (a T5 text encoder reads the prompt). A decoder turns the tokens back into sound you can hear. At each step the model scores 2,048 possible next tokens, keeps the top 250, and picks one at random. That randomness is why the same prompt gives a different song every time — and why you always get two versions to choose from.

Your words arrive in three channels: style text (genre, mood, instruments, era, vocal character) through a text encoder; structural tags like [Verse], [Chorus] and [Bridge] as arrangement instructions; and lyrics through a separate alignment process that fits words to melody and rhythm. There are two model families behind it all: autoregressive — token by token, like a chatbot (Suno, MusicGen) — and diffusion — start from noise and clean it into sound (Stable Audio). Suno's current generation is reported to run a hybrid of both.

Our verdict: WATCHING. AI music is real and already flooding streaming — but almost nobody listens to most of it, and the law is still being written.

The flood nobody listens to

Spotify had to remove over 75 million spammy tracks in a single 12-month period leading up to September 25, 2025. It now labels AI artist profiles with specific AI persona badges. Spotify isn't just curating music anymore — it is fighting a daily war against automated noise.

Deezer's June 2026 data shows the scale: about 90,000 fully AI-generated tracks are uploaded every single day — more than 50% of all daily uploads, up from 28% in September 2025. But fully AI tracks account for only 1–3% of total streams. And of those AI streams in 2025, up to 85% were fraudulent — played by bots in streaming farms trying to steal royalty payouts. Machines making music for other machines to listen to.

Apple Music reported in April 2026 that more than a third of its monthly deliveries are now fully AI — yet AI tracks are under 0.5% of listening time. Spotify told ABC News in July 2026 that "well under 1 per cent" of consumption goes to fully AI-generated tracks.

And your ears cannot help. A Deezer survey of 9,000 people in eight countries found 97% could not tell fully AI-made songs from human ones in a blind test — while 80% of those same people want such music labelled. We can't hear the difference at all, but we really want to know who — or what — is singing.

Suno's race: v3 to v6

This didn't take decades. Version 3 dropped in March 2024, proving the concept — the first tool where anyone could type one sentence and get a full song with vocals in under a minute. By September 23, 2025, version 5 arrived with massively improved quality. Then v5.5 landed on March 26, 2026, bringing Voices (record your own singing voice, verified against a spoken phrase) and Custom Models. Studio 2.0 followed on August 13, 2026 — a browser-based studio with MIDI on a timeline. Then the major milestone: v6 launched on September 9, 2026, built in partnership with Warner Music Group, BMG and Believe, with songs up to 8 minutes long.

The rivals: Google's Lyria 3 Pro (best for emotional, cinematic tracks), ElevenLabs Music, Stable Audio 2.5 (best control over timing and structure). Udio — once Suno's main rival — has had downloads paused since August 9, 2026 during a rights transition.

What it costs — and the real bottleneck

Suno Free is $0 — 50 credits a day, v6-mini only, no downloads, no commercial rights. Pro is $10 a month ($8 billed yearly) — 2,500 credits a month, 20 song downloads a month, v6 and v6-wild, Voices, commercial rights on songs made while subscribed. Premier is $30 a month ($24 billed yearly) — 10,000 credits a month, 60 downloads, Suno Studio 2.0.

The real bottleneck is not the credits. It is the strict monthly download limits — and since September 3, 2026, the caps apply even to songs made before the caps. You can generate variations all day, but you can't export everything. You have to curate. A full song, typically 2–4 minutes, renders in about 30–60 seconds, and every generation gives you two takes to pick from.

The 5-part prompt formula

Here it is — the practical payoff. To command the machine, define five things: Genre + Mood + Instruments + Vocal style + Structure.

Example: "Dreamy indie-folk, acoustic guitar and soft piano, 90 BPM, warm female vocal, verse-chorus-verse — a song about the last train home."

If you use this exact formula, you guide the text encoder to a highly specific cluster. You get a structured, professional-quality track instead of a random jumble of sounds.

Do you own the song? The lawsuits

But does that text prompt make you the legal author? The US Copyright Office was clear on January 29, 2025: prompts alone do not make you the author. Typing a sentence doesn't cross the threshold of human authorship. You can generate a masterpiece, but legally, you didn't write it. The machine did. And a machine cannot hold a copyright.

The industry is totally split. Warner settled in November 2025, opting for partnership — that deal is why v6 exists. But Universal Music Group and Sony went to war, suing Suno on September 18, 2026. Shortly after, Reservoir sued on October 21, 2026 over 1,050 copyrighted compositions. The EU AI Act's transparency rules (Article 50) have applied since August 2, 2026 — AI systems that generate audio must mark their output as artificially generated in a machine-readable way.

The full verdict: the mechanism is real — tokens, transformers and diffusion — and the songs are genuinely good enough to fool 97% of listeners in a blind test. It is already useful for creators, ads, demos and anyone who wants a song that never existed. But most AI songs are never heard, the courts have not settled who owns what, and the winning tool just rebuilt itself on licensed music. Use it to make things; do not build a career on pretending it is human.

Type one sentence. Get a song. Just remember — the band that proved it worked doesn't exist.

Paragon AI MAG — one video and one podcast episode every day, every fact checked against sources. Watch the video on YouTube and listen to the full conversation on the Paragon AI MAG podcast.


Watch the video: https://youtu.be/LrLgGAsU-dg

Listen to the podcast: https://redcircle.com/shows/cbce814a-d451-4913-a015-fe05b4d0e059

Originally published on Substack: https://paragonaimag.substack.com/p/how-ai-music-generation-works

Top comments (0)