AI Music Generators Compared: Which Fits a Video Soundtrack?
Compare Mureka, ElevenLabs Music, MiniMax Music 3.0, Lyria 3.5 and Cassette AI for video soundtracks, plus ways to fix repetitive, overly literal AI lyrics.

Which AI music generator should you use for a video soundtrack?
It depends on what the soundtrack needs. If you want a full song with vocals, or you want to build on a melody you hummed, Mureka fits. If you only need background music with no singing, Cassette AI is instrumentals only. ElevenLabs Music, MiniMax Music 3.0 and Google's Lyria 3.5 are the other options. One creator who tested these for a game trailer found Lyria 3.5 fit best. ElevenLabs struggled with the style he wanted. That is one test, so treat it as a starting point.
TL;DR
- No single model wins every job. The creator who compared them said he couldn't name one "better" model, only the one that suited his project.
- Lyria 3.5 by Google was the favorite for a high-energy game trailer soundtrack, while ElevenLabs Music missed on that particular style.
- Mureka offers full vocal songs and reference tracks, so you can hum or tap a rough melody and build on it.
- Cassette AI makes instrumentals only, which suits dialogue-heavy videos where singing would compete with the voiceover.
- Plan to generate many versions. Expect a lot of misses before a keeper, so budget time for volume.
- AI lyrics fail in predictable ways: they repeat themselves and restate your prompt literally. You can fix both with better instructions and a human edit.
What were these models tested on?
The comparison came from a creator making a 30 to 90 second animated promo for a game he is building, a football game set in a fantasy world. He made the visuals and music inside one platform, Genspark Media Studio, which bundles many image, video, voice and music models in a single workspace. The soundtrack was meant to be, in his words, an absolute banger to get the blood pumping.
That context matters. A trailer wants energy, a strong hook and a clear genre. A corporate explainer or a training video wants something quieter that stays out of the narrator's way. A model that wins on one can lose on the other.
How do the five music models compare?
Here is what the source supports about each. Where nothing specific was stated, the table says so rather than guessing.
| Model | What was noted | Best fit |
|---|---|---|
| Mureka | Full vocal songs. Accepts an uploaded reference track, such as a hummed or tapped melody, and builds on it. | Songs with singing, or projects that start from your own rough idea |
| ElevenLabs Music | Described as a great tool that has produced good music in the past. Did not do well with the specific style wanted this time. | Worth testing, but check it against your genre |
| MiniMax Music 3.0 | Available in the same studio. No strengths were detailed. | Include in your test batch |
| Lyria 3.5 (Google) | The creator's favorite for this trailer. | High-energy trailer-style music, based on one test |
| Cassette AI | Instrumentals only, no vocals. | Background music under narration or dialogue |
The pattern is simple. Pick by job first (vocals or no vocals, from scratch or from a melody), then pick by taste.
Why did Lyria 3.5 win in this test?
The creator did not give a technical reason. He said which one he liked best for his specific style, and that ElevenLabs did not handle that style well even though it had made great music for him before.
That is a useful reminder for anyone choosing a tool. Music models have genres they handle well and genres they don't. A model that nails a lo-fi study beat may fumble an orchestral-rock trailer. Don't read "Lyria won" as "Lyria is better." Read it as "Lyria matched this brief."
How do you pick the right model for your project?
Ask three questions in order.
- Does the video need vocals? If not, an instrumental-only model like Cassette AI removes a source of problems, because there are no lyrics to go wrong. If yes, you need a model that sings, such as Mureka.
- Do you already have a melody or a feel in mind? Mureka's reference track option lets you hum, tap a rhythm on a desk, or upload a rough idea, and the song is built on top of it. That gives you more control than describing a sound in words.
- What genre and mood? Run the same brief through several models and compare. This is where a model like Lyria might beat one you expected to win.
How many tries does AI music take?
Many. The creator's main advice was not to expect a hit on the first try. He recommended working by volume: test different styles and different lyrics, and treat the first outputs as sketches.
A practical routine:
- Write one short brief: genre, tempo feel, mood, length, and whether you want vocals.
- Run it through three or four models.
- Listen to each with your video playing, not on its own. A track that sounds great alone can clash with your cuts.
- Keep the best two, then vary the style words and regenerate.
- Only then fine-tune lyrics, if you are using vocals.
Many platforms charge credits per generation, so a small test batch before a big one saves money.
Why do AI lyrics sound repetitive and too literal?
Large language models, the chatbots that write most AI lyrics, have habits. The creator mentioned that he knows shortcuts these models take, though the details were not captured in the material this article draws on. Two habits are well known from general experience:
- Repetition. The same rhyme pairs, the same stock phrases, and a chorus that restates the verse. Models gravitate toward the most common phrasing, and in song lyrics the most common phrasing is a cliché.
- Over-literal writing. Ask for a song about a football game and you get lines that say "football," "field," and "game" over and over. The model treats your prompt as a checklist to recite.
How do you get better AI lyrics?
These are general techniques, not a quote from any one creator. They work because they push the model away from its default phrasing.
- Describe a feeling and a scene, not a topic. "Two rivals at the end of a long night, nobody wants to go home" beats "a song about a rivalry."
- Ban the stock words. List the clichés you see in the first draft and tell the model not to use them again.
- Ask for specific details. One concrete image per verse (a taped wrist, a cracked scoreboard) gives the song texture.
- Ask for subtext. Tell the model the song should imply the theme and never name it.
- Vary the structure. Request different line lengths and an unrhymed line or two. Perfect rhyme on every line is a giveaway.
- Write the hook yourself. The chorus is the part people remember. A human-written chorus with AI verses often reads better than the reverse.
- Edit by ear. Read the lyrics aloud. Cut any line you would be embarrassed to say.
Here is a prompt you can copy and adapt:
Write lyrics for a [genre] song for a [length] video about [subject].
Do not name the subject directly. Imply it through scenes and images.
Mood: [two or three feelings].
Rules:
- Each verse needs one concrete, unusual image.
- No repeated rhyme pairs across verses.
- Avoid these words and phrases: [list clichés from your last draft].
- Vary line length. Rhyme only where it sounds natural.
- Keep the chorus to four lines or fewer, simple enough to sing along to.
Give me two different versions and tell me which lines are weakest.
The last line helps. Asking the model to critique its own draft often surfaces the clichés before you spend credits generating audio.
Who does this change things for?
- Small teams and solo creators who used to pay for stock music or a composer can now generate custom tracks for promos, explainers and social clips in an afternoon.
- Marketers and internal comms staff can score a training video or a product demo without licensing hassles, though you should check each platform's terms for commercial use before publishing.
- Anyone using vocals should budget time for lyric editing. The audio can be polished while the words still read like a template.
What you can do: pick one real project, run a three-model test with the routine above, and save your winning prompts so the next soundtrack starts from a working recipe.
Could you build your own soundtrack request and review tracker?
If several people in your company ask for music or video clips, the mess is rarely the generating. It is the tracking: who asked, which brief, which versions were tried, which one was approved. A small internal app can hold a request form, a version list with notes and an approval status, instead of a spreadsheet and a chat thread. With Remy, you describe that tool in plain English and it builds and deploys it as a full-stack app. Remy's catalog includes nearly 400 AI models, music among them, so the same app could also draft lyrics from your brief.
Frequently Asked Questions
Which AI music model is best for lyrics and vocals?
Of the models in this comparison, Mureka was the one highlighted for full vocal songs and for building on a reference track. The other models may also sing, but the source only confirmed that Cassette AI does not. Test your lyrics on two or three vocal-capable models before committing.
Can I use a hummed melody to guide the song?
Yes, with Mureka. You can upload a reference track, even a hummed melody or a rhythm tapped on a desk, and the song is built on top of it. This gives you more control than a text description alone.
Is Lyria 3.5 better than ElevenLabs Music?
Not in general. In one creator's test, Lyria 3.5 fit a trailer-style brief better, and ElevenLabs didn't handle that style well, though it has produced great music for him before. Run your own brief through both.
Do I need vocals for a video soundtrack?
Often not. If your video has a narrator or dialogue, an instrumental track avoids competing voices and skips the lyric-writing problem. Cassette AI is instrumentals only, which suits that case.
Why do AI lyrics repeat the same words?
Language models tend to choose the most common phrasing, which in songs means clichés and recycled rhymes. Give the model a scene and a feeling instead of a topic, ban the stock words, and edit the result yourself.
Useful for
How we source, write, and correct our articles. Editorial standards
Related articles

How Does ReShoot 4D Change a Video's Camera Angle With AI?

Fish Audio Live 3 Pricing: What Voice Agents Cost Per Minute

How Much Does an AI Short Film Cost? The $102,000 Breakdown
