How do you turn speech or a voice recording into a song with AI?

It is 5:40 in the morning and your friend Marco sends a voice memo from the bakery he just sank everything into. Fifty-three seconds of oven noise, a laugh, and a line you want to keep forever: "I'm not sure. I'm here though." His opening day is Saturday, and you want to give him a song made out of that memo.

So you type "speech to song AI" into a search bar and get a page of tools that all promise the same thing in different words. They do not do the same thing. Some keep the recording and wrap music around it. Some keep the voice and try to make it sing. Some keep only the words. Pick the wrong one and you burn an evening on a result that is technically "your memo as a song" and sounds like none of what you wanted.

This page sorts the routes, says what each tool on each route does according to its own maker, and shows the route that works most reliably for a gift: the words, turned into a real lyric.

Three jobs hide behind "speech to song"

Before choosing a tool, decide which part of the memo you are trying to keep. That one decision picks the route.

If you already have finished lyrics and only need something to sing them, the AI singer comparison is your page: it compares what sings and how to pick the voice. This page is for the step before that, when what you hold is a recording of someone talking and you have not decided what part of it survives into the song.

Why talking does not turn into singing by itself

Speech has pitch. Every sentence rises and falls. What it lacks is stable notes: a spoken syllable slides through a range of pitches instead of landing on one, and the rhythm follows breath and meaning, not a beat.

There is a well-documented exception, and it explains a lot about which routes work. The psychologist Diana Deutsch at UC San Diego noticed that a short spoken phrase from one of her own recordings started to sound sung once it was looped. With Rachael Lapidis and Trevor Henthorn she tested it and presented the results in 2008: a phrase heard once sounded like speech, the same phrase repeated several times started to sound like song. The effect needed exact repetition. When the repeats were transposed, or the syllables shuffled, listeners kept hearing speech. It is called the speech-to-song illusion.

Two things follow for anyone making a song from a memo:

• A spoken recording, played once, gives a music model very little melody to hold on to. Any tool that "keeps" it is either inventing a tune or leaving it as speech over a beat.
• Repetition is what turns words into something the ear treats as music. That is the logic of a chant or a looped hook, and it is the reason the strongest move on route 3 is to make the person's own sentence the line that repeats.

Route 1: upload the memo and build music around it

Suno lets you bring your own audio. On the Create page there is an audio upload button next to Create; you can import a file or record a clip on the spot, and the upload is saved to your library. From there Suno's help centre names Extend (continue the track from a point you choose) and Cover (keep the melody and move the track into a different style). Suno's own announcement of Covers lists a simple voice memo as a possible starting point.

Three things to know before you try it with speech:

• Length limits depend on the plan and have changed. At the time of writing, Suno's help articles disagree with each other: one says free accounts can upload up to 60 seconds and Pro or Premier up to 8 minutes, an older one says 6 to 60 seconds with up to 120 for paid plans. Check the limit inside the app before trimming.
• You confirm you own the material. The upload flow asks you to agree that you own what you are uploading. A friend's voice memo is your friend's recording. Ask them, and for a gift, asking is half the fun anyway.
• Suno does not document what happens to a clip with no tune in it. Cover is described as keeping the melody. A spoken memo does not have one in the musical sense, so expect the model to decide for you. Budget several attempts and listen to each before you commit.

When this route fits: you want the memo itself to be audible, the laugh, the oven noise, the "I'm here though", with music growing out of it. When it does not: you want a song with verses and a chorus about the memo. That is route 3.

For the craft of steering an upload (keeping the Style field short, stopping Suno from sneaking in a key change, extending from the right second) there is a separate walkthrough on remixing your own audio in Suno. Everything there applies to a voice memo, with the caveat above about melody.

One practical trick that sits between route 1 and route 3: make the song by route 3, then put the original memo in front of it as a spoken intro. Any audio editor can do that cut, no AI involved, and you get both the real voice and a song that has a shape.

Route 2: make the voice itself sing

This is what people usually picture when they type "voice to song AI": the same voice, now singing. There are two families of tools, and neither turns a spoken take into a sung one by itself.

Voice conversion swaps the timbre of a recording for another voice. Kits describes its conversion system as studio-quality singing voice conversion and says it carries over the input's pitch, dynamics and small details like breathiness. That is the point: it follows the performance it is given. Feed it a spoken sentence and the pitch line it follows is a spoken pitch line. The useful workflow is the other way round: someone sings or hums the melody, and conversion puts a different voice on it.

Vocal synthesizers start from a score instead of a recording. Dreamtonics describes Synthesizer V Studio 2 Pro as software where you enter notes and lyrics, choose a voice and shape the expression. At the time of writing Dreamtonics sells it as a one-time purchase, not a subscription. You get control; you also have to write the melody yourself, note by note.

So route 2 answers "how do I get a particular voice singing" and assumes you already have a melody and words. It does not answer "how do I get a song out of this memo".

A note on whose voice: putting a real person's voice on a song is a consent question before it is a technical one, and in some countries a legal one as well. This is not legal advice; our disclaimer and the article on who owns AI-generated music set out what we can and cannot say about rights.

Route 3: keep the words, write a lyric, then have it sung

This is the route that most reliably ends in a song with verses, a chorus and the details that make it about one person. It has four steps, and only the third one runs on sunomarket.

Step 1: get the transcript

You need the memo as text. A few options, checked on the makers' own pages:

• Apple's Voice Memos app shows a transcript of a recording on supported devices. Apple lists English, Spanish, Portuguese, Italian, French, German, Japanese, Korean and Chinese for this feature, and notes it is not available in every country or region.
• Telegram Premium lists voice-to-text conversion for voice messages among its features, which matters if the memo arrived as a Telegram voice note.
• Whisper, OpenAI's open-source speech recognition model, is released under the MIT licence, installs locally and supports dozens of languages. It suits anyone comfortable with a command line or a desktop app built on it.

If what you hold is a song rather than speech, the problem is different (whose words, do they exist as text already), and getting the words out of audio covers that case.

Step 2: pick out the lines that could only belong to this person

A transcript is not a lyric. It is mostly filler, false starts and half-sentences. Here is Marco's memo, as a transcriber would hand it back:

ok so it's 5:40, I'm at the bakery, the oven's been on since four,
um, the sign guy still hasn't come so the window just says BAKE,
no RY, ha. anyway. I sold the car. I sold the car for flour basically.
my mom keeps asking if I'm sure. I'm not sure. I'm here though.
first loaf's out. it's ugly. I'm keeping it. ok. bye.

Read it with a pen and circle only what nobody else's memo would contain. Here that is: the oven on since four, the half-finished sign that reads BAKE, the car sold for flour, the mother asking if he is sure, his answer, and the ugly first loaf he is keeping. Six details. "It's 5:40" and "anyway" go. The answer "I'm not sure. I'm here though" is the one to protect almost word for word, because it is the hook. In the brief below it becomes "not sure, but I'm here": two short beats a chorus can carry, with the meaning untouched.

Step 3: turn those details into a brief

The sunomarket lyric generator takes a short form, not a transcript. The theme field takes up to 800 characters, but pasting the whole transcript would waste it: every extra detail competes with the six that should reach the chorus. For Marco:

Theme:     Gift song for my friend Marco opening his bakery. Oven on since
           4am, the half-hung sign reads BAKE, he sold his car to buy flour,
           his mom keeps asking if he's sure, he says "not sure, but I'm
           here". The first loaf came out ugly and he kept it.
POV:       second
Language:  en
Structure: classic
Hook:      title-line
Vocal:     male
Emotion:   hope
Age:       any

That theme is 243 characters, under a third of the cap, and all six details made it in. Two choices are worth explaining. POV second because the song talks to Marco, which is what a gift song usually does. Hook title-line because his own sentence should become the title and come back in every chorus, which is the speech-to-song idea moved into a lyric: the line that repeats unchanged is the one that starts to feel like music.

What a lyric built from that brief looks like

Here is a lyric written by hand for this article to show the shape. Every fact in it comes from the theme field above; the rest is the kind of imagery a lyric adds around those facts. Nothing comes from the parts of the transcript that were cut:

[Verse 1]
The oven's been burning since four in the morning
The sign on the window says BAKE and no more
You traded the car for a mountain of flour
And opened the bakery door

[Pre-Chorus]
Your mother keeps asking if you're really sure

[Chorus]
Not sure, but I'm here
Hands in the dough, I'm here
Nobody promised the morning would come out right
Not sure, but I'm here

[Verse 2]
The first loaf came out ugly and crooked
Not one you would ever put out
Anybody else would have thrown it away
You kept it, and that's what it's about

[Chorus]
Not sure, but I'm here
Hands in the dough, I'm here
Nobody promised the morning would come out right
Not sure, but I'm here

[Bridge]
Half of a sign and a car's worth of flour
That's a whole lot of here

[Final Chorus]
Not sure, but I'm here
Not sure, but I'm here

The generator returns a title and a full lyric with section tags like these, ready to paste into Suno's Lyrics field. It does not sing. Audio is a separate step in a tool you choose, and for most people that is Suno.

Step 4: get it sung

Paste the lyric into the Lyrics field of the singing tool you picked, write a Style prompt for the sound, and generate. Suno is a different company: sunomarket is an independent site, not affiliated with Suno or with any other company named on this page, and we do not resell or control Suno. If the voice keeps drifting between takes, locking a vocal persona explains how to hold it steady, and song structure in Suno covers how the section tags are read.

If you already have an instrumental and a paid Suno plan, Suno's help centre describes Add Vocals, a beta feature for Pro and Premier users: you open an uploaded track or a generated instrumental, enter lyrics, and Suno layers a sung vocal on it. That is another way to finish route 3.

Which route fits which reader

• "I want to hear the actual memo inside the song." Route 1, or route 3 with the memo cut in as an intro. The intro version gives you more control.
• "I want it sung in a specific voice." Route 2, but you will need a melody, which means someone sings a guide take or you write notes in a vocal synth. Route 3 still gives you the words.
• "I want a real song about this person, by Saturday." Route 3. It is the only one where the details of the memo end up in verses and a chorus instead of in the background.
• "I have a recording of someone singing and want the words." Not this page: that is transcription of a song, see the audio-to-lyrics article linked above.

What route 3 costs on our side

The lyric step is the only paid part on sunomarket, and the numbers are simple. One generation on the SM1 model costs 20 credits, which is $0.20 at 100 credits to the dollar. SM2 and SM3 cost 80 credits ($0.80); they run an extra editing pass in which a detector flags weak lines and a repair step rewrites them. New accounts start with 0 credits, and there is no free trial of the lyric generator.

To get credits you need a plan: $5 a month gives 500 credits (25 SM1 lyrics), with unspent plan credits cleared at renewal; $19 once gives 1 500 credits that never expire (75 SM1 lyrics). It is not unlimited generation. Details are on the pricing page. The singing tool is billed separately by whoever makes it.

Two limits worth knowing before you paste anything. A generated lyric is not guaranteed to be unique: lyrics are non-exclusive and can overlap between users. And we claim no authorship over what you make with the generator; the disclaimer spells out the rest.

Do it tonight, in this order

1. Open the memo and get a transcript, whichever tool your phone or app already has.
2. Circle the details nobody else's memo would contain. Aim for four to six, plus the one sentence you would hate to lose.
3. Write them into the theme field (the cap is 800 characters, a few sentences is enough), set POV to second if the song speaks to the person, and set the hook to title-line so their sentence comes back in every chorus.
4. Generate the lyric at /generators, read it against the memo, and check every detail is one they would recognise.
5. Take the lyric to your singing tool, and if the real voice matters, cut the memo in front of the song as the intro.

The memo already has the hardest part of a gift song in it: something only that person would say. Every route above is a different way of keeping that one line alive.