The Kivelo jingle, generated with the Suno V6 API in English, Spanish, French, German, Japanese and Mandarin.
Brands that advertise in several countries rarely want six different songs. They want one recognisable jingle, sung in each market’s language, with the brand name pronounced the same way everywhere. Traditionally that means booking singers in six languages, which is expensive and slow. AI music models promise to do it in minutes, but most demos only show English. The obvious question for a marketing team is whether the other languages are good enough to air.
We tested that with a concrete brief. We invented a green smoothie brand called Kivelo, wrote a short jingle, translated it into five more languages, and generated every version with the Suno V6 API using the same musical style. Then we measured how accurately each version was sung, using speech recognition as an impartial listener. We also checked how the made-up brand name survived in each language, compared two Suno versions, produced 15- and 30-second ad cuts, and normalised them to streaming loudness. In total we made 11 Suno requests for 42 cents.
All six language versions can be played in this article, and a three-minute walkthrough video at the end shows the charts with short excerpts.
Why measure with speech recognition
A native speaker is the best judge of a sung lyric, but six native speakers are not always available, and opinions vary. Automatic speech recognition gives a repeatable number. We transcribed every take with Whisper, set to the correct language, and compared the transcript with our lyric sheet. For English, Spanish, French and German we compared words. For Japanese and Mandarin, which do not separate words with spaces, we compared characters. We report “lyric accuracy”, the share of lyric-sheet words or characters that were heard correctly. We left extra repeated words out of that number, because singers often repeat a line, and we report them separately as insertions.
This method has limits, and we will point them out as we go. Whisper makes its own mistakes, especially with homophones and invented names, so the numbers are an upper bound on Suno’s errors, not an exact measure of them. But because every language went through the same process, the comparison between languages is fair.
import time, requests
BASE = "https://api.apimart.ai/v1/music"
H = {"Authorization": "Bearer YOUR_KEY"}
STYLE = ("upbeat ukulele pop jingle, handclaps, glockenspiel, cheerful female vocal, "
"110 bpm, radio advert jingle for a green smoothie brand")
def generate(lyrics, version="v6"):
tid = requests.post(f"{BASE}/generations", headers=H, json={
"custom": True, "version": version, "title": "Kivelo Jingle",
"prompt": lyrics, "style": STYLE, "vocal_gender": "female", "duration": 30,
}).json()["data"][0]["task_id"]
while True:
d = requests.get(f"{BASE}/tasks/{tid}", headers=H).json()["data"]
if d["status"] in ("completed", "failed"):
return d
time.sleep(6)
jingles = {lang: generate(text) for lang, text in LYRICS.items()}Test 1: writing the lyrics
We kept the jingle short, four lines with the brand name twice, which is typical for a radio or pre-roll spot. The translations are adaptations rather than word-for-word copies, so each version rhymes and scans naturally in its own language. Here are all six, with English back-translations:
English
Morning light and a bright new day / Kivelo, fresh in every way
Shake it up and take it slow / Taste the green, Kivelo!
Spanish
Luz del sol y un nuevo día / Kivelo, pura alegría
Agítalo y sin prisa voy / Sabor a verde, Kivelo hoy
(Sunlight and a new day / Kivelo, pure joy / Shake it, I’m in no hurry / A taste of green, Kivelo today)
French
Lumière du matin, un jour nouveau / Kivelo, frais et si beau
Secoue-le, prends ton temps / Goût de vert, Kivelo, maintenant
(Morning light, a new day / Kivelo, fresh and so beautiful / Shake it, take your time / A taste of green, Kivelo, now)
German
Morgenlicht, ein neuer Tag / Kivelo, so wie ich’s mag
Schüttel ihn und nimm dir Zeit / Grüner Genuss, Kivelo, bereit
(Morning light, a new day / Kivelo, just how I like it / Shake it and take your time / Green delight, Kivelo, ready)
Japanese
朝の光 新しい日 / キヴェロ いつもフレッシュ
ふって ゆっくり / みどりの味 キヴェロ
(Morning light, a new day / Kivelo, always fresh / Shake it, slowly / The taste of green, Kivelo)
Mandarin
清晨的阳光 美好的一天 / 基维洛 新鲜每一天
摇一摇 慢慢来 / 绿色好味道 基维洛
(Morning sunlight, a beautiful day / Kivelo, fresh every day / Shake it, take it slow / Green, good taste, Kivelo)
For Japanese we wrote the brand in katakana as キヴェロ (ki-ve-ro), and for Mandarin we used the transliteration 基维洛 (jī-wéi-luò), which is how a brand would typically appear in each market. Each version was sent with section tags, [Verse] and [Chorus], the same style description, a female vocal and a target duration of 30 seconds.
Test 2: speed, length and consistency
All eight requests completed in 29 to 37 seconds and cost $0.05 each for two takes, so the six-language set cost 30 cents. The style held well: 15 of the 16 descriptions Suno returned mentioned ukulele pop, and most also named the handclaps and glockenspiel we asked for. The voice did not hold as well. Although every request asked for a female vocal, the descriptions ranged from “high, light, breathy” to “low, dark register” and “gravelly rasp”, and the takes sound like different singers. For a brand that wants one recognisable voice across markets, that is a real limitation, and Suno’s persona feature, which reuses a singer from an existing track, is the tool to test next.
We asked for 30 seconds. All 16 takes landed between 28.8 and 30.0 seconds.
The duration control was impressively precise. In an earlier test with 90-second background music we saw one take run more than 50 percent long, but for these 30-second jingles every one of the 16 takes was between 28.8 and 30.0 seconds. For advertising, where slots are fixed, that precision matters.
The first take in each language. Same style prompt, six lyric sheets.
English (v6-wild, take 1).
Spanish (v6, take 2).
Test 3: lyric accuracy in six languages
Share of the lyric sheet heard correctly, two v6 takes per language.
English scored highest at 91.3 percent on both takes. Japanese followed at 88.6 percent on both, then German at 84.2 and 89.5, Spanish at 80.0 and 90.0, and Mandarin at 87.5 and 75.0. French was clearly lowest at 71.4 and 76.2 percent. Two things stand out. First, every language except French reached at least 84 percent on at least one take, so with two takes per request there was almost always a usable version. Second, the difference between takes was often larger than the difference between languages: Spanish moved by ten points and Mandarin by twelve between take 1 and take 2. The practical advice is simple: always check both takes, because the better one is often much better.
German (v6, take 2): 89.5 percent lyric accuracy.
Test 4: what the errors actually were
Numbers alone can mislead, so we read every transcript and grouped the errors.
Examples of differences between the lyric sheet and what Whisper heard.
French suffered the most, and part of the reason is French itself. “Goût de vert” (taste of green) was transcribed as “goût de verre” (taste of glass), but “vert” and “verre” are pronounced identically, so that was a transcription limit, not a singing mistake. “Secoue-le” (shake it), on the other hand, was heard as “ce couleur” and “sa couleur” in both takes, which suggests the phrase really was sung unclearly. In Mandarin, the opening “清晨” (qīngchén, early morning) was heard as “青春” (qīngchūn, youth) in both takes, and “摇一摇” (yáo, shake) became “咬一咬” (yǎo, bite) in one take, a tone slip that a Mandarin listener would notice. In Japanese, the lyrics were nearly perfect apart from, in one take, the brand name, which we discuss next. German had only small slips, such as “Schüttel ihn” heard as “Schütte ihn”.
We also normalised harmless differences before scoring. Whisper writes Mandarin in traditional characters, so we converted them to the simplified characters we used. Japanese was written with the kanji 緑 and 降って instead of our kana みどり and ふって, which are the same words read aloud (although 降って is actually the wrong meaning, “fall” instead of “shake”). We counted these spelling choices as correct, because a listener would hear exactly the same sound.
French (v6, take 2): listen for “Secoue-le” in the second half.
Test 5: did the brand name survive?
For a jingle, the brand name is the most important word. “Kivelo” is made up, so speech recognition cannot simply look it up, and the way it is transcribed tells us how it sounded.
How Whisper wrote the brand name in each take.
The results were the reverse of what we expected. German got it exactly right in both takes, transcribed as “Kivelo” and “Kivelow”. Mandarin produced the exact characters 基维洛 in one take. Spanish gave “quibelo” or “qui velo” in every take, and French gave “qui vélo” and “qui veut l’eau”. Those are phonetically correct: in Spanish, b and v sound the same, and “qui vélo” is literally “ki-ve-lo” written as French words. Japanese consistently produced キメロ (ki-me-ro) instead of キヴェロ (ki-ve-ro), so the unusual “ve” sound was not clear enough.
English was the worst. In the two v6 takes the brand name was heard as “keep it all” and “keep it low”, and in v6-wild as “kibble-o” and “keep the look”. “Kivelo” sounds very close to common English phrases, and when it is sung quickly the listener, human or machine, hears the phrase. That is a valuable lesson for brand naming as much as for AI music: test how your name sounds when sung in each market before you print it on packaging. If the name must stay, writing it phonetically in the lyrics, for example “Kee-veh-low”, is a common trick to force the intended pronunciation.
Japanese (v6, take 1): the brand is sung closer to “ki-me-ro”.
Test 6: v6 versus v6-wild
Suno V6 offers several versions, and v6-wild is described as the more creative one. We generated the English and Spanish jingles with both.
Substitutions (S), deletions (D) and insertions (I) for each take.
“Creative” turned out to be an accurate label. The v6 Spanish takes added no extra words at all. Both v6-wild Spanish takes repeated the chorus, adding 13 and 11 words, and one English v6-wild take changed the lyrics, singing lines that Whisper heard as “shake, shake, tap” and “heave a down”, with 13 inserted words and only 69.6 percent accuracy. The other English v6-wild take was as accurate as v6 (91.3 percent). For a jingle where the client has approved every word, v6 is the safer choice. For brainstorming variations of a hook, v6-wild’s improvisation might be exactly what you want.
Every jingle in this test, in all six languages and both versions, came from the same Suno V6 API call with only the lyric sheet changed, which is what makes this kind of multi-market production fast to script and easy to repeat after a lyric change.
Test 7: cutting a 15-second and a 30-second ad
Broadcast and pre-roll slots usually come in 15- and 30-second lengths. The 30-second version was already the right length, so we only needed a short ending. For the 15-second version we used Suno’s crop operation on the best Spanish take:
crop: task_id = Spanish v6 take 2, start_s 0, end_s 15
The crop cost $0.008, took 15 seconds, and returned a file of 15.01 seconds, only 13.5 milliseconds longer than requested, which is the normal padding of the MP3 format. We also tried Suno’s fade-out operation on the 15-second cut (at $0.008 each). For the final ads we applied the closing fade in our own mastering step instead, so that the fade length matched the ad specification to the millisecond, a step most studios keep in their own hands anyway.
Test 8: loudness for streaming and radio
Streaming platforms and many ad networks normalise to around –14 LUFS and ask for true peaks no higher than –1 dBTP, to prevent distortion after encoding. We measured all 16 raw takes.
The final 15- and 30-second ad cuts after normalisation and limiting.
The raw takes were already close to the loudness target, between –13.5 and –15.7 LUFS. But 12 of the 16 takes had true peaks above –1 dBTP, and one English take peaked at +0.4, which means it would clip in some players. After a gain adjustment and a gentle limiter, our final cuts measured –14.3 LUFS for the Spanish 15-second version, –14.2 LUFS for the German 30-second version and –14.3 LUFS for an English 30-second version, with peaks between –1.4 and –2.4 dBTP. That is ready for most ad platforms. The lesson is that Suno’s output is loud enough, but not peak-safe, so always run a final limiter before delivery.
Mandarin (v6, take 1): 87.5 percent character accuracy, with the brand name transcribed exactly.
A measurement lesson: speech recognition invents text
One more finding is about the measuring tool, not the music. In three cases Whisper added a sentence at the end of a transcript that was never sung: “Subtítulos realizados por la comunidad de Amara.org” in Spanish and its French equivalent. This is a known artefact. Whisper was partly trained on subtitled video, and when it hears music without speech, it sometimes reproduces subtitle credits. We removed these lines before scoring, and it is a reminder that any automated quality check needs a human to look at the outliers.
How we would run this for a real campaign
Based on these results, this is the workflow we would use for a real multi-market jingle. It keeps the speed of the API and adds human checks where the measurements showed risk.
First, have a native copywriter adapt the lyrics in each market, not just translate them, and ask them to mark the brand name with a phonetic hint where the local pronunciation is not obvious. Second, generate each language with v6, and request two or three times, which gives four to six takes per market for 10 to 15 cents. Third, run every take through speech recognition automatically and discard anything below about 80 percent lyric accuracy, as well as any take whose length is outside the ad slot. In our test that filter would have removed the weaker Mandarin take and the improvised English v6-wild take, and it would have rejected both French takes, a clear signal to rewrite the French lyric before spending any more time on it.
Fourth, check the brand name separately. A simple rule works well: if the transcript contains the brand spelled correctly or phonetically, the take passes, and if it contains an ordinary word or phrase in its place, such as “keep it low”, it goes to a human. Fifth, send the two best takes per market to a native speaker for a final listen, with the transcript next to them so they can focus on the words the machine flagged. Finally, crop to the slot length, apply your own fade and limiter, and deliver at –14 LUFS with true peaks below –1 dBTP. In our test, the automated part of this process took under ten minutes for all six languages. The human review is where the time goes, and that is exactly where it should go.
The walkthrough video
The video below shows the charts with short excerpts of each language version.
Walkthrough: one jingle in six languages with the Suno V6 API (about 3 minutes).
What it cost
| Step | Requests | Time per request | Cost |
|---|---|---|---|
| Jingle in six languages, v6 (12 takes) | 6 | 29–36 s | $0.30 |
| English and Spanish, v6-wild (4 takes) | 2 | 36–37 s | $0.10 |
| Crop to 15 seconds | 1 | 15 s | $0.008 |
| Fade-out tests | 2 | 15 s | $0.016 |
| Total | 11 | $0.42 |
Six language versions of a jingle, two takes each, plus alternative versions and ad cuts, cost 42 cents. The Whisper transcriptions we used for measurement are not included, because they are part of the test, not of the production. Even if you add a native-speaker review for each market, which we would recommend before airing, the cost is a small fraction of a traditional multi-language session.
Checklist for multi-language jingles
- Adapt the lyrics, do not translate them word for word. Rhyme and rhythm in each language matter more than literal meaning.
- Always compare both takes. Accuracy differed by up to 12 points between two takes of the same request.
- Use v6 for approved lyrics and v6-wild for ideas. v6-wild repeated and changed lines.
- Test the brand name in every language. Ours was heard correctly in German, phonetically in Spanish and French, and as an English phrase in English.
- Write tricky names phonetically if needed. Unusual sounds such as “ve” in Japanese may need help.
- Expect French and tonal languages to need extra review. Homophones and tones caused most errors.
- Limit the peaks before delivery. 12 of 16 raw takes exceeded –1 dBTP.
- Treat automatic transcripts as a filter, not a verdict. Whisper invents subtitle credits and confuses homophones.
The bottom line
For 42 cents, the Suno V6 API produced a jingle in six languages, each in the same musical style, all within 1.2 seconds of the requested length. Speech recognition heard between 71 and 91 percent of the lyrics correctly, with most languages above 84 percent on their better take. The weak points were predictable: French phrases with ambiguous sounds, Mandarin tones, an unusual syllable in Japanese, and a brand name that sounded too much like English words. None of those are reasons to avoid AI jingles, but all of them are reasons to have a native speaker listen before a campaign goes live. If you want to build a similar multi-market workflow, or combine music with voice, image and video models under one key, you can find them all on apimart.ai.