AI Music API: Lyria 3 Pro vs ElevenLabs Music vs Sonilo, and Which Are Safe for Commercial Use

Every brand team I talk to wants custom music now. Nobody wants to be the person whose product launch gets a copyright claim. Those two facts are why the AI music API question has changed. A year ago it was "which one sounds best". Today it is "which one can I actually put in a paid ad".

The backdrop matters. In June 2024 the major labels sued Suno and Udio, and the settlements that followed in late 2025 turned on licensing. So when a vendor says its model is "commercial-safe", I want to know what that phrase is standing on: what the model was trained on, who owns the output, what the plan you are on actually allows, and who pays if something goes wrong.

I ran the same briefs through the three music models on Segmind: Google DeepMind's Lyria 3 Pro, ElevenLabs Music and Sonilo. That came to 18 API calls covering an ad bed, a brand jingle with my own lyrics, a two-minute podcast bed, a video score, a Hindi jingle and a copyright guardrail test. The whole run billed $2.17. Below is what came back, measured rather than eyeballed, and then what each vendor's terms actually say about commercial use.

The three models at a glance

All three are on Segmind as plain REST endpoints: Lyria 3 Pro, ElevenLabs Music and Sonilo as Sonilo Text to Audio, Video to Audio and Video to Video. One key, one bill.

How I tested

Same brief to every model wherever the model could take it. Every clip is embedded below. Duration and format came from ffmpeg. Loudness came from EBU R128, as integrated LUFS and true peak. Whether the lyrics were sung as written came from Whisper transcripts. Whether a requested structure happened came from tracking low-frequency energy (kick and bass) in 2.5-second windows. Costs are what Segmind's request history says each call billed.

Test 1: a 30-second ad bed for a marketing team

This is the bread-and-butter job: an agency needs a fresh instrumental under a 30-second spot, and needs it to be exactly 30 seconds.

ElevenLabs and Sonilo both came back at 30 seconds to within two hundredths of a second. Lyria 3 Pro returned a 60.3-second track. That was not a one-off: all five times I asked Lyria for something under a minute (10 or 30 seconds), it came back between 58 and 68 seconds. The model page says length is "influenced" by the prompt, and in practice there seems to be a floor around one minute. For an ad, that means an editor trims it, and the trim point will not land on the musical ending.

The loudness numbers tell a second story. Lyria's track measured -12.5 LUFS with a true peak of +0.8 dBFS. That is over full scale, so you will want a limiter before it goes to a broadcast spec. ElevenLabs measured -14.4 LUFS at -0.7 dBFS and Sonilo -14.0 LUFS at -0.9 dBFS, both sitting where streaming platforms normalise. Every Lyria file in this test peaked above 0 dBFS.

On arrangement, the spectrograms differ. Sonilo's bed has a break just after the top, a lift in the middle and a breakdown before a final hit, which is the shape of an ad cue. ElevenLabs stays at an almost constant level (loudness range 1.2 LU), which is fine under a voiceover and dull on its own. Lyria built risers, but placed them inside its own one-minute structure, not at the 0:20 I asked for.

import requests

r = requests.post(
    "https://api.segmind.com/v1/elevenlabs-music",
    headers={"x-api-key": "YOUR_API_KEY"},
    json={
        "prompt": "Upbeat 30-second instrumental bed for a running-shoe ad ...",
        "music_length_ms": 30000,
        "model_id": "music_v2_5",
        "force_instrumental": True,
        "output_format": "auto",
    },
)
open("ad-bed.mp3", "wb").write(r.content)

Test 2: a brand jingle with your own lyrics

Jingles are where commercial risk gets personal, because legal approves the exact words. I wrote a two-line verse and a chorus for a made-up coffee brand and checked what got sung.

ElevenLabs sang my verse and chorus, two separate transcripts found nothing added, and it stopped at 30 seconds. Lyria sang my lines word for word too, then kept going. It wrote a second verse ("Sip by sip and you start to smile, stay for a while") and a new chorus line ("Warmest coffee, shining so light") that nobody approved. For a demo that is generous. For a brand, it is copy that has not been through legal. Sonilo ignored the lyrics completely and returned an instrumental, which is exactly what its documentation promises.

So I tried to fix Lyria with its own tools: timestamps on every section, plus an explicit instruction to sing only the lyrics as written.

That worked on the words. The verse landed between 0:00 and 0:12 and the chorus between 0:12 and 0:27, and the transcript matched my script exactly with nothing added. The file still ran 58 seconds, though. The song ends at about 37.5 seconds, followed by roughly five seconds of true digital silence and then a ten-second instrumental reprise. The lesson for Lyria: timestamps are the lever for content, and you still need to trim.

Test 3: a two-minute podcast bed with a structure

Production houses and podcast networks need longer beds with a shape: a quiet intro under the host's open, a groove, and an outro that clears space. I asked for drums and bass to enter at 0:15 and drop out at 1:30. Lyria got the timestamps in the prompt. ElevenLabs got a three-chunk composition_plan with the same timings. Sonilo got three music_segments starting at 0, 15 and 90 seconds.

At two minutes, Lyria hit the length almost exactly at 118.99 seconds, so its length floor clearly sits below the length of a full song. On structure, Lyria and ElevenLabs behaved almost identically. Both opened sparse with no bass, brought the bass in at about 10 seconds rather than 15, and dropped it at about 105 seconds rather than 90. That is the right shape, early in and late out. Sonilo had bass from the first second, kept the groove going to the very end and faded only in the final two seconds. The only trace of its 90-second segment was a brief dip near 92 seconds. If the outro matters, Sonilo needed the most editing here.

This is also where pricing flips. Lyria charges $0.10 per song whatever the length. ElevenLabs and Sonilo charge per second, so two minutes cost $0.375 each, nearly four times as much.

Test 4: scoring a finished video

For film and video teams, the real job is often "put music under this cut". Sonilo is the only one of the three that reads a video. I took a silent 10-second headphone spot from my Kling 4 tests and sent it to Sonilo Video to Video in music_and_sfx mode. Then I gave Lyria two frames from the same clip as reference images, and gave ElevenLabs a 10-second text brief.

Sonilo returned the clip as an MP4 with the soundtrack already muxed in, matched to the frame (10.04 seconds). The spectrogram shows the score following the edit. A soft pad sits under the macro shots, and a rhythmic pulse arrives at about 3.4 seconds, right as the camera pulls back to reveal the person. Nobody told it where that moment was. Billing was exactly $0.0225 per second of video.

Lyria returned 65.5 seconds for a 10-second brief, so it is a library track to cut down, not a score. ElevenLabs delivered exactly 10.03 seconds for three cents and came in quieter (-19.5 LUFS), which leaves headroom for a voiceover. Neither of them knows where the cut is. If your input is a finished edit, Sonilo is doing a different and more useful job.

Test 5: a Hindi jingle

A lot of Segmind's customers ship regional ads, so I wrote a short Hindi tea-brand jingle in Devanagari, with the musical direction in English.

Both sang in Hindi. A whisper-large-v3 transcript of the ElevenLabs track recovered my verse and chorus. This time ElevenLabs also added Hindi lines of its own ("दिल को दे सुकून", "नई सुबह नया जोश"), something it did not do in English.

The guardrail test: naming an artist

Last, I sent all three a prompt asking for a country-pop song "in the style of" a famous living singer. This is the request a careless user will make, and what happens next is part of what commercial-safe means.

• Lyria 3 Pro blocked it in 1.8 seconds through the provider's content moderation. Billed $0.00.
• ElevenLabs Music rejected it in 2.3 seconds with the reason copyrighted_material_detected, and sent back a rewritten prompt with the name removed. Billed $0.00.
• Sonilo generated a 30-second instrumental and billed $0.094. It can't imitate a voice, because it doesn't sing, but it did not stop the request. I am not publishing that clip.

Two out of three refusing a named artist, for free, is a real safety net. With Sonilo the guardrail is your own prompt hygiene.

Which are safe for commercial use?

Here is what each vendor's own published terms say, checked on 1 November 2026. I'm not your lawyer, and terms change, so follow the links before you rely on any of it.

Sources: Gemini API terms, Google's Lyria 3 Pro post, Google Cloud indemnified services, Eleven Music terms, Eleven Music launch, Sonilo licensing, Sonilo terms.

Three things jump out. First, none of the three indemnifies you for music output, so "commercial-safe" means "lower risk", not "someone else's risk". Second, the training-data stories are not equal. Sonilo names a licensed catalog. ElevenLabs names licensing partners. Google describes rights it holds without naming a catalog. Third, the plan you are on matters as much as the model. ElevenLabs and Sonilo both put film, TV and broadcast behind their top tiers, and Sonilo fixes your rights at the plan you were on when the track was generated.

One more thing to know if you call these through an API gateway like Segmind: your rights flow through the gateway's agreement with the provider, not through the vendor's consumer plans. For ads, social and apps, that is rarely a question. For a theatrical release, a TV spot or a big game launch, get the answer in writing before you ship.

My read, by job:

• Instrumental beds for ads, apps and social: Sonilo has the cleanest training-data story and exact durations, and nothing it makes can be mistaken for a singer.
• Jingles and songs with approved lyrics: ElevenLabs Music. It held to my English lyrics and length, and it actively refuses artist names. Check the plan you are on before anything goes to TV.
• Long-form beds and volume: Lyria 3 Pro, at $0.10 per song regardless of length, with SynthID marking every track as AI. Trim it, limit it, and remember it is a Preview model.
• Scoring a finished cut: Sonilo Video to Video. It is the only one that watches the footage.

What it costs on Segmind

All of these are the published Segmind rates, and every call in this test billed them exactly. ElevenLabs bills the returned duration to the hundredth of a second (30.02 seconds cost $0.0938125). Sonilo's video endpoints cost $0.01125 per second for music or effects alone and $0.0225 per second for both. Lyria 3, the 30-second clip version, is $0.05 per generation. Failed and refused calls billed nothing.

Honest assessment

ElevenLabs Music was the most predictable: exact lengths, my English lyrics as written, and a refusal that came with a usable rewrite. It was also the fastest, at 6 to 12 seconds per call. Its weakness was flat dynamics on the ad bed, and it padded the Hindi lyrics. Lyria 3 Pro gave the most music per dollar and followed a timestamped structure well, but it ignores short durations, adds lyrics unless told not to, and clips above 0 dBFS. Sonilo nailed durations and video sync and has the strongest licensing narrative, but it can't sing, followed my segments loosely, and didn't stop a named-artist prompt.

FAQ

Which AI music API is safest for commercial use?

Sonilo has the clearest training-data claim (licensed catalogs, including Shutterstock), and ElevenLabs Music has label licensing partners plus strict prompt rules. None of the three offers indemnity, so check your plan's media rights before broadcast use.

Can Lyria 3 Pro make a 30-second track?

Not reliably. In my tests every request under a minute returned 58 to 68 seconds. Ask for the length you need, then trim. At $0.10 flat per song, the extra length costs nothing.

Does ElevenLabs Music sing my exact lyrics?

In English it sang my verse and chorus as written and added nothing. In Hindi it sang my lines and then added its own. Always transcribe or listen before a jingle ships.

Can Sonilo generate vocals?

No. Sonilo generates instrumental music and sound effects only. It ignores lyrics in the prompt and returns an instrumental.

How much does an AI music API call cost on Segmind?

Lyria 3 Pro is $0.10 per song. ElevenLabs Music and Sonilo Text to Audio are $0.003125 per second, so $0.1875 a minute. Sonilo video scoring is $0.01125 to $0.0225 per second.

Will these models copy a famous artist?

Lyria 3 Pro and ElevenLabs Music refused a named-artist prompt in my test, at no charge. Sonilo generated an instrumental anyway, so keep artist names out of your prompts.

Wrapping up

For ads I would reach for Sonilo, for jingles ElevenLabs Music, for volume and long beds Lyria 3 Pro, and for scoring a finished cut Sonilo Video to Video. All of them are a single API call on Segmind, billed per song or per second, with no setup. Try ElevenLabs Music, Lyria 3 Pro or Sonilo with your own brief, and listen to what comes back before legal does.