Atmee.aiDocs
Guides

Avatar 101: Negative voice examples

The voice samples that make an avatar sound flat or wrong, each with a clip to listen to and the fix. Most problems are in how the sample is read, not in the microphone.

Your avatar speaks the way your voice sample does. This page lists only what goes wrong, with a short clip of each problem, so you can rule it out. Anything not on it is fine.

A good sample is 10 to 60 seconds of one person speaking clearly; 15 to 30 seconds is plenty. Formats and size limits are in Create an avatar.

The clips are not a template

Every clip is an invented voice with one problem added, so that the problem is the only thing that differs. Speak in your own way: any accent, language, pace or style works.

Why the voice matters

The voice sample drives more than how the avatar sounds:

  • Lip sync. The avatar's mouth moves to its voice. A flat, quiet or muffled sample is a common reason the lips barely move or fall out of time with the words.
  • How the avatar comes across. Its energy, warmth and character on every call come from your sample. A tired or monotone read gives an avatar that sounds bored, however good the portrait is.
  • Generation. Atmee builds the avatar's talking clips from your portrait and voice together. A noisy or mixed sample can make the takes stiff or wrong, and Atmee refuses a sample shorter than 10 seconds or longer than 60.

How each problem is marked

  • Won't work: Atmee refuses the sample.
  • Likely to fail: the avatar usually sounds or moves wrong.
  • May work: try it, and listen to the result.

Delivery

The most common problem isn't the microphone: it's a sample read without life. A voice that rises and falls, stresses the words that matter and sounds like it means what it says gives an avatar that does the same, with lips that move clearly. Read flat, and every answer comes out flat.

  • Likely to failRead in a flat, even voice
    The avatar speaks the way your sample does. With little rise and fall, the voice sounds bored and the lips barely move.Fix: Talk the way you would to a friend: let your voice rise on questions, fall at the end of a sentence, and lean on the words that matter.
  • Likely to failTired and listless
    Low energy carries straight into the voice, so every answer sounds like the end of a long day.Fix: Record when you feel awake, sit or stand up straight, and speak a little louder and livelier than feels natural.
  • Likely to failEach word read on its own
    Reading word by word, or reading out a list, gives the voice a choppy, robotic rhythm.Fix: Speak in whole sentences, as if you were explaining something. Tell a short story or talk freely instead of reading items out.

Recording

  • Won't workShorter than 10 seconds
    There is too little speech to learn a voice from, and Atmee refuses a sample under 10 seconds.Fix: Record 15 to 30 seconds; anything from 10 to 60 seconds works.
  • Won't workLonger than 60 seconds
    Atmee refuses a sample over 60 seconds.Fix: Trim it to your best 15 to 30 seconds.
  • Likely to failSomeone else speaks in part of the file
    The voice is learned from everyone in the file, so the avatar can drift between two people.Fix: Cut the other person's part out and keep only your own speech.
  • Likely to failTwo people talking at the same time
    Overlapping voices can't be pulled apart, so the voice comes out muddled.Fix: Record again on your own, in a quiet room.
  • Likely to failMusic under the voice
    The music is learned as part of your voice and comes back as hum and warble.Fix: Record with no music playing.
  • Likely to failBackground noise
    Chatter, traffic or a fan is learned along with your voice, and loud noise fails Atmee's check.Fix: Record somewhere quiet, with windows closed and fans off.
  • Likely to failToo quiet
    A quiet recording is mostly room sound. Atmee may refuse it, and the voice comes out thin.Fix: Hold the microphone or phone closer, about a hand's width from your mouth.
  • Likely to failToo loud and distorted
    The crackle is learned too, so it comes back in every word the avatar says.Fix: Move the microphone a little further away or turn the input level down, then record again.
  • Likely to failLong silences
    Long pauses leave too little speech, and silence at the start or end can fail Atmee's check.Fix: Trim the silence at the start and end, and keep pauses short.
  • May workMuffled or far from the microphone
    A muffled or distant voice loses the detail the voice needs, so the avatar sounds dull.Fix: Speak towards the microphone from close by, without covering it.
  • May workRoom echo
    The echo becomes part of the voice, so the avatar always sounds like it is in a bathroom.Fix: Record in a room with soft furnishings, like a bedroom, not a bare kitchen or hallway.
  • May workPhone-call quality
    A recorded call is thin and narrow, and the voice keeps that sound.Fix: Record with your phone's voice memo app, not over a call.

If the result isn't right

Replace the voice sample with a new one. A new voice builds a new version of the avatar, and you can switch back to the earlier one.

How to optimize your avatar says which fix to try first for what you see and hear.

Next steps