Music teachers: isolate the take before you clone
Clone consent. Isolate vocals. Do not clone a bed.
Clone consent. Isolate vocals. Do not clone a bed. A lesson take is almost never a voice. It is a voice sitting on a piano, a metronome, a hallway, and last week's backing track. A clone trained on that mix is a clone of the mix. Every generated count-in will arrive with the same upright, whether you wanted it or not.
Isolate the voice before you clone it is the order. Voice-clone consent is the client clause is who may be in the sample. This page is the studio version of both: do not clone a recital tape.
Consent is the person, not the lesson
A cloned voice is a person, reproduced, saying words they have not yet said. If that person is you, the teacher, write the grant: uses (lesson recaps, ads, practice tracks), languages, and what happens when you stop teaching. A verbal "sure" on a Zoom is not a grant.
If that person is a student, stop. A recital is not a style reference. Do not generate a child's voice for ads because the take was "just for Instagram." Clone the teacher who signed, or pick a stock voice.
Do not hide this in a prompt library. A saved instruction that says "use Maya's clone" is a workflow, not a right. AI voice cloning is the tool after the clause exists. A "sound like this singer" brief is a living-person impression. Refuse it.
The bed is not a voice
Cloning has no "ignore the piano" knob. The model learns a speaker from a slice of audio. Room tone, a metronome, a second player — those become part of the timbre.
Run isolation on the sample if the take has piano or a printed bed, a Zoom hallway, studio murmur, or a recap recorded over last week's track. Do not expect isolation to fix clipping, two talkers, or a thin distant voice in a tiled practice room. Far-mic'd speech is mostly room. The clone will be thin forever.
Isolate vocals from a track is the general-purpose pass on any mixed audio_url. It is not the Suno stem split. That other tool needs a prior generate_music id. A lesson tape is an upload. Use isolate.
Headphones test: mute everything except the person. If what is left is not the teacher you will ship, re-record in a dead room. If it is, isolate, then clone.
Isolate, then clone, then teach with it
Capture for the clone, not for the recital: close mic, dead room, 10–30 minutes, spoken cues as well as sung vowels. If the capture is not that room, isolate and hunt piano ghosts on the esses. Clone from the isolated sample only. Name it. Generate a paragraph the sample never said. Park the noisy original — it is not a backup clone.
Cloning is billed once per sample. Re-cloning because the first sample was a mixed recital is how you pay twice for a worse voice. Reuse a good voice_id.
A teacher sends a new line recorded in the car. Do not clone again. Isolate that line if you need that performance, or generate it from the existing clone. Weekly dirty clones are how the library fills with "Studio Recital Mix" nobody should have saved.
FAQ
Can I clone first and isolate the generated speech later?
You can isolate output. You cannot un-learn a piano the model already ate. Clean the sample.
Is a 30-second clean clip better than ten minutes of recital tape?
Yes. A short clean sample clones a limited range; a long dirty sample clones a limited range and a Yamaha. Get clean, then get longer.
The student said I could use the recital on Instagram. Does that cover a clone?
No. Using a recording and generating new speech are different acts. If there is no written clone grant, do not clone.
Do I isolate every lesson upload, or only the clone sample?
Isolate the clone sample, always. Isolate a lesson upload when you need that performance without the bed. One clean voice_id is the library.