Craft
Why your AI voice clone repeats words
If the same stray words appear at the start of every generated section, inspect the recording used to clone the voice before rewriting the script.
5 min read ·
We saw a Hindi generation repeat part of its reference recording at the start of each output chunk. The typed script was different each time, yet the repeated words were the same. That pattern pointed to the voice reference, not the script.
The clue: repetition at every chunk boundary
A long script is generated in smaller sections and joined afterward. If a phrase appears once, it may be a script or model error. If it appears at the beginning of every section, the shared input to those sections is a better place to look. In our case, that shared input was the cloned voice's reference audio and its transcript.
Why the reference transcript matters
Mayantar's Indian-language path uses a reference recording together with the words spoken in it. The model keeps only the first 12 seconds of reference audio. If the transcript still includes words spoken after that point, the model sees words it cannot match to the audio. It may then speak those missing words as if they belonged at the start of the new output.
A practical check before you retry
- Listen to the first few seconds of the generated audio. Write down the words that repeat.
- Listen to the original voice recording. Check whether those words occur near its end or after the first 12 seconds.
- Record a fresh, clear clip that finishes a complete sentence comfortably before the 12-second stop. Do not try to squeeze in extra words at the buzzer.
- Generate a short test sentence in the same language as the new recording. If it starts cleanly, then try the full script.
If it still repeats
Try another voice on the same short script. If only one voice repeats, its reference is still the likely cause. If multiple voices repeat on one script, inspect the script and its language setting. A Hindi reference generally works better for Hindi output than an English reference; the reference recording guide explains how to make a clean one.
This diagnosis is specific to a transcript-conditioned cloning path. Other models can repeat for other reasons, including difficult text or generation instability, so a new clip is a test rather than a universal cure.