VoxCPM2: finding a voice for narration and characters

Choose where the voice comes from

You may have a finished script without a voice, or a recording of yourself that you want to use for new narration. These need different preparation. VoxCPM2 offers speech generation, description-based voice design and reference-based cloning.

Without a recording, begin by describing the timbre and delivery. With a recording, check that the speech is clear before previewing a clone. For dialogue, save and test the characters separately so two similar voices do not become difficult to distinguish.

Read the specifications in context

Upstream lists 30 languages, voice design, controllable cloning and native 48kHz output. Language count describes coverage; sample rate describes an output specification. Neither tells you whether a particular narration has the right pronunciation, emphasis or pauses.

Controllable cloning adds expression guidance to a reference voice. A hurried delivery, background sound or a strong accent in the reference can still affect results. Start with one specific change, such as “keep the timbre, speak slightly more slowly, as if explaining a step,” and listen for the difference.

Use a test that reveals problems

A greeting alone will not tell you how a voice handles your script. Include a number, a question and a longer sentence. Here is an original test passage:

We need to finish chapter twelve today. The train leaves at three twenty. Do we still have time? Take a breath, and bring the blue notebook from the desk.

Keep the words fixed while choosing the voice, then keep the voice fixed while adjusting delivery. Check numbers, question intonation and whether the character fits the context. Follow the test with the opening paragraph of your actual project.

Save the voice in Sonano

1. Prepare VoxCPM 2 in the Model Library and open the voice design or cloning workflow.
2. Describe the voice or choose an authorized recording, then generate a short sample.
3. Listen, save a suitable voice to the Voice Library, and select the matching model, output language and voice in the speech workflow.
4. Preview a short exchange for dialogue, or a chapter for an audiobook.

The dialogue and narration samples in Sonano’s home page showcase use VoxCPM2. They are finished examples with fixed scripts and voices. Test your own language, material and Mac separately.

When to compare another model

If the timbre feels wrong, revisit the voice source first. If the voice fits but individual words fail, check the script and pronunciation. When comparing models, keep the test passage fixed and consider clarity, character fit and the number of corrections.

Continue with the Qwen3-TTS variant guide, or check the VoxCPM 2 overview for current app options.