Similarity depends on what the clone was built from. "Clone of the original voice" at upload is assembled automatically from the speech in your video, and we do not control the quality of that speech. Your own voice created from a clean recording in "Voices" is almost always closer.
How a clone from the video is made #
- We separate the voice from music and noise.
- Using the transcript timestamps we cut out the speaker's lines. Parts where someone else talks at the same time are dropped.
- The longest lines are joined into a sample of up to 60 seconds, and the voice is created from it.
At least 10 seconds of clean speech are needed. With less, no clone is created and the speaker is voiced with a standard library voice.
A clone from the video is always set up to speak the target language without an accent. The speech is cleaner, but similarity drops, see Accent in the translation.
What hurts a clone from the video #
- Music and noise that could not be fully separated. The clone repeats everything it hears in the sample.
- Several people talking at once or interrupting each other.
- Little speech: a short video or a speaker who rarely talks. 20 seconds is worse than 60.
- Echo, a phone recording, a distant microphone, shouting or whispering.
- Two different people merged into one "speaker". This happens when fewer speakers were set than are in the video, or when recognition mixed them up, see How many speakers to set.
How to get the closest voice #
- Create your own voice in "Voices" from a clean recording: one person, no music, 30 seconds or longer. See How to record a sample for your voice. Available on Pro and Business.
- Leave "Remove accent" off if similarity matters more than native pronunciation.
- When uploading, pick that voice for the speaker under "My voices".
If the translation is already done, change the speaker's voice in the editor on the "Speaker voices" tab. All of the speaker's phrases are re-voiced, and minutes are charged by character count, see What uses minutes.
If similarity is still not good enough #
Write to support with a link to the translation. We will listen to the sample that came out of your video and tell you what exactly gets in the way.
How voice cloning works and how services differ is covered in a separate guide.