Prepare a better reference recording for voice cloning

Choose a passage with one speaker

A reference recording is the voice material your model hears. Overlapping speakers, music and strong room echoes make it harder to isolate the person you want. Start with a clear passage containing only the target speaker.

If you can record again, keep the microphone position and speaking level steady, and avoid fan noise, typing and handling sounds. Listen back for clipping and cut-off words. Follow your model and app requirements for duration and format; joining recordings from many different rooms does not necessarily help.

Match the transcript to the recording

For a cloning route that requires reference text, write what the recording actually says. If the speaker says “See you Friday,” do not supply “We will meet on Friday afternoon.” A polished paraphrase no longer matches the audio.

Automatic transcription can give you a draft. Check names, numbers, foreign words and fillers while listening. If you trim either end of the recording, update the text too. The standard Qwen3-TTS cloning example pairs reference audio with its transcript.

Test with something new

Ask the model to read a short passage that was not in the reference recording. Include a name, a number, a question and a longer sentence from the kind of work you plan to make. Listen for voice identity, pronunciation and pauses.

1. Keep the model, language and test text fixed while comparing two reference recordings.
2. Choose a reference, then adjust generation settings in small steps.
3. Save the recording with its matching transcript and reuse the test passage for later checks.

A disappointing take can come from the source material, text, language or generation settings. Short tests help you narrow down the cause before you generate a long script.

Save the voice for your project

In Sonano, save the cloned voice to the Voice Library with a recognizable name, then select it for narration or a dialogue character. For a cast, preview a complete exchange to check that the speakers remain easy to distinguish.

Test cross-language output separately. A similar timbre does not guarantee the accent and pronunciation you want. Use your own recording or material you have permission to use, within the agreed scope.

Continue with voice cloning and the Voice Library, or read about the Qwen3-TTS variants.