AI voice cloning lets you turn a short recording of your own voice into a voice you can type with. Once the clone exists, every script you paste into Text to Speech comes back in your voice - no microphone, no retakes, no booking studio time.
The quality of the result depends far more on the sample you give the model than on any setting you change later. This guide shows you how to record a good sample, create the clone in Pynokio and fix the problems people most often run into.
What is AI voice cloning?
A voice cloning model listens to a reference recording and learns what makes that voice recognisable: pitch, timbre, accent, rhythm and the way you shape vowels and consonants. When you then give it new text, it speaks that text with the same vocal identity.
Pynokio runs cloning on its own speech engine. You upload or record a sample, the engine uses up to 60 seconds of it as the reference, and the finished voice appears in your voice list next to the ready-made library voices. From there you can use it in Text to Speech exactly like any other voice.
What you need before you start
- Your own voice, or written permission. You may only clone a voice you own or have specific, written authority to clone. More on this below.
- A quiet room. Soft furnishings help: a room with curtains, a sofa or a wardrobe full of clothes sounds far better than an empty kitchen.
- Any decent microphone. A USB microphone or a good headset is ideal, but a modern phone held at a steady distance works too.
- A short script to read. Two or three paragraphs of natural text, in the language you want the clone to speak.
Step 1: Record a clean voice sample
You can record directly in the browser or upload an audio file (up to 20 MB). Recordings and uploads can be up to 70 seconds long, and the engine uses up to 60 seconds of them. Aim for 30 to 60 seconds of continuous speech - long enough to capture your range, short enough to stay consistent.
Recording tips that make the biggest difference
- Keep the distance constant. About a hand's width from the microphone, slightly off-axis so your breath does not hit it directly.
- Speak the way you want the clone to speak. If you want calm narration, read calmly. The clone copies your delivery, not just your timbre.
- No music, no other voices. Background music, TV or a second speaker confuse the model about which voice to learn.
- Avoid heavy processing. Skip effects, reverb or aggressive noise gates. A plain, slightly quiet recording is better than a "polished" one.
- Read naturally, not word by word. Full sentences with normal pauses teach the model your rhythm.
Tip: If your only recording has hum, hiss or room noise, turn on Remove background noise when you create the clone. Use it only for noisy recordings - it can add up to two minutes of processing, and clean samples are kept unchanged anyway.
Step 2: Add the transcript
The engine works best when it knows exactly what you said in the sample. Pynokio can transcribe the sample for you automatically, or you can paste the text you read. Check the transcript for missing or misheard words - an accurate transcript helps the model line up sounds with letters, which shows up later as cleaner pronunciation.
Step 3: Confirm consent
Before the clone is created, you confirm that the voice is yours or that you hold written authority from the voice owner, and that you will disclose synthetic speech where people could mistake it for a real recording. This is not a formality: cloning someone without permission is prohibited by our Voice Policy and, in many countries, by law. Read our guide to voice cloning consent and the law if you are cloning a voice for a client or a colleague.
Step 4: Create the clone and test it
Give the voice a name you will recognise, pick its language and enter a short preview sentence. When the clone is ready, listen to the preview with fresh ears:
- Does it sound like you on a good day, or like you on a bad phone line? The second usually means the sample needs improving.
- Is the accent right? If not, check that the voice language matches the language of your sample.
- Are there odd breaths, clicks or room echo? Those come straight from the sample.
If something is off, it is almost always faster to record a better sample and create a new clone than to fight it with settings.
Step 5: Use your cloned voice in Text to Speech
Open Text to Speech, choose your new voice and paste your script. Three settings shape the delivery:
| Setting | What it does | When to change it |
|---|---|---|
| Speed | Makes the voice faster or slower | Tight video timing, or a slower, explanatory pace |
| Stability | Lower values add variation and expression, higher values keep the delivery even | Raise it for long narration, lower it for lively, conversational reads |
| Clarity | How closely the output sticks to the character of the reference voice | Raise it if the clone drifts away from your voice |
Download the result as MP3 for everyday use or WAV if it is going into a video or audio editor. For more on writing scripts that sound natural, see our text to speech guide.
Troubleshooting common problems
The clone sounds robotic or flat
The sample was probably read too carefully. Record again, reading the way you would explain something to a friend, and lower Stability slightly when you generate.
There is noise or echo in every result
The model reproduces the room it heard. Record in a softer space, move closer to the microphone, or enable background noise removal for the sample.
Some words are mispronounced
Names, brands and abbreviations are the usual suspects. Spell them the way they sound in your script (a surname like "Nguyen" can be written "Win") and write numbers out as words.
The accent changes between generations
Make sure the voice language and the script language match, and keep Stability in the middle of the range for long texts.
Voice cloning vs. voice design
Cloning copies a real voice. Voice Design creates a brand-new voice from a written description, such as "a warm, low male voice with a light Irish accent, speaking slowly". Use cloning when the voice has to be yours; use design when you need a distinctive character without a real person behind it. Our voice design guide shows how to write good descriptions.
Is a cloned voice detectable?
Yes, by design. Every speech file generated in Pynokio carries an inaudible watermark and comes with a machine-readable provenance record. Anyone can check a file for free with the Audio Detector. This protects you too: if someone claims a recording of "you" is real, the watermark helps show where synthetic audio came from. Read how AI audio watermarking works.
Key takeaways
- The sample decides the quality: 30-60 seconds of clean, natural speech in a quiet room.
- An accurate transcript and the right language setting improve pronunciation.
- Only clone your own voice or a voice you have written permission to use.
- Fine-tune delivery with Speed, Stability and Clarity in Text to Speech.
FAQ
How long should my voice sample be?
Between 30 and 60 seconds of continuous, clean speech works best. Pynokio accepts samples up to 70 seconds and uses up to 60 seconds as the reference.
Can I clone someone else's voice?
Only with their specific, written permission covering AI voice cloning and how the voice will be used. Cloning public figures, colleagues, customers or family members without that authority is not allowed.
Which languages can my cloned voice speak?
Text to Speech supports 30 languages. A clone sounds most natural in the language of its sample, so record the sample in the language you plan to use most.
Can I delete my voice clone?
Yes. Deleting the voice removes it from your account together with its samples and preview files, and revokes its consent record.
Do I need special equipment?
No. A quiet room matters more than an expensive microphone. A phone or headset at a steady distance is enough for a good clone.
Ready to hear yourself? Create your first voice clone in a few minutes - your sample stays private and is never published.
