The short version
- Length: 5 to 15 seconds. abcVoiceover accepts 3 seconds to 10 minutes, and Fast draft mode needs at least 5.
- One speaker, with no music, no other voices and no background noise.
- A quiet, soft room: carpets, curtains and furniture absorb echo; bare walls and tiles bounce it.
- Steady distance: about a hand's width to a forearm's length (15 to 30 cm, 6 to 12 inches) from the mic, the whole time.
- The right performance: speak the way you want the voice to sound in its lines.
- WAV format, or press Record in browser in abcVoiceover, which saves WAV for you.
Why the sample matters so much
abcVoiceover uses zero-shot voice cloning. The model doesn't spend hours learning a voice; it listens to your one clip and imitates it for every line it reads. That's why a new voice is ready the moment its clip is uploaded, and it's also why the clip matters so much. The model copies the timbre, accent and pace you want, but it copies the room echo, the laptop fan and the half-hearted delivery too.
Step by step
- Pick the room. A small room with soft furnishings is best: a bedroom, or even a wardrobe full of clothes. Avoid kitchens and bathrooms. Turn off fans, air conditioning and notifications, and close the window.
- Set up the microphone. A USB or podcast mic is ideal, but a decent headset or a phone held steady works. Sit slightly off to the side of the mic rather than straight into it, which avoids popping on p and b sounds, and keep the same distance throughout.
- Choose what to read. A few natural sentences in the voice's usual register, not a list of words. Mix short and long sentences so the model hears how the voice moves. There are two examples below.
- Perform it. Read the way the character or narrator will usually speak. If the villain snarls, snarl; if the course narrator is warm and steady, be warm and steady. Don't whisper or shout unless that's how every line should sound.
- Record. In a project's voices, press Record in browser and allow the microphone, or record in any audio app and export a WAV file to upload. Leave a moment of quiet before you start speaking, but trim long silences.
- Listen back on headphones. Check for hum, hiss, echo, clicks and distortion (a harsh, crackly sound when you got loud). If you hear any, record again; it's ten seconds well spent.
- Test it. Generate two or three lines with different moods. If they all sound right, you're done. If not, the table of fixes below usually has the answer.
Two passages to read
For a character voice:
I've walked this road a hundred times, but tonight it feels different. The wind is colder, the trees are quieter, and somewhere up ahead there's a light on that shouldn't be. Well. Let's find out who's home.
For a narrator or instructor:
Welcome to the course. Over the next few minutes, we'll look at three simple habits that make a big difference, and by the end you'll be able to try each of them yourself.
Common problems and how to fix them
- Takes sound echoey or distant: the room is in the clip. Record somewhere softer, closer to the mic.
- Takes have hiss or hum: background noise is in the clip. Turn off the noise source, or turn the mic's gain down and move closer.
- Takes sound flat: the sample was flat. Record it with more energy, or pick the Expressive style and raise emotion intensity in Fine-tune.
- Takes rush or drag: the model follows the clip's pace. Record at the pace you want, or lower guidance strength in Fine-tune to loosen it.
- An accent in another language: expected. A voice speaking a language other than the one in its clip keeps some of the clip's accent.
- Fast draft won't run: it needs a clip of at least 5 seconds, and it's English only.
Whose voice can you record?
Your own, or that of someone who has clearly agreed to their voice being cloned for your project. The terms of service require it, and our guide to using AI voices responsibly explains how to ask and what to write down.