Voice design, voice cloning and speech generation: how they fit together

Separate choosing a voice from reading a script

You want to narrate a story, but the model list includes VoiceDesign, Base and CustomVoice alongside cloning and speech generation. It helps to split the task: find the voice, then ask that voice to read the script.

Design and cloning previews also produce audio, but their purpose is to establish the voice. Production generation is where you check whether the current script has been read accurately and completely.

Voice design starts with a description

For a fictional narrator without a reference recording, describe audible characteristics. For example:

An adult man with a low, warm and clear voice, speaking at a steady pace. He is telling a story to one person across a table, with sentences that settle naturally.

The design model reads sample text according to that description. Listen, revise the description if needed, and save a suitable result to the Voice Library. Include a longer sentence, number or character name from the kind of work you plan to make.

Sonano’s current design engines include Qwen3-TTS VoiceDesign, VoxCPM2 and OmniVoice. They handle descriptions and languages differently. OmniVoice upstream notes that design training uses Chinese and English data; evaluate other-language design results separately.

Voice cloning starts with a recording

Use a recording of yourself or a speaker you have permission to use, with a transcript matching what was actually spoken. Save it as a reusable voice. A speech model can then refer to those voice characteristics when reading new text.

Sonano’s current cloning import stores managed reference audio, its transcript and language information. It does not train a separate model for every voice. The selected speech model processes those reference inputs when generating a new passage.

The sample may convey accent, rhythm and recording conditions alongside timbre. Music, overlapping speech or heavy echo can affect the result. See the reference recording guide for preparation.

Speech generation reads the new script

Once you have a voice, supply the production script to a speech model. The new script does not have to repeat the reference recording. Reference text explains what the sample says; the production script contains what you want spoken this time.

Reuse the voice across passages, assign voices to dialogue characters, or generate an audiobook chapter by chapter with its project-level model, language and voice. Listening and revising a line or chapter are part of production.

You can also start with an existing voice. Distinguish a model’s native presets from Sonano’s supplied reference voices: they use different generation paths.

What the models do in Sonano

The table reflects Sonano’s current integration. Qwen3-TTS VoiceDesign, Base and CustomVoice are different models within a family, with different jobs.

Reference-based speech uses a voice sample to read new words: this is where cloning capability participates in generation. Qwen also documents a workflow that designs a short sample with VoiceDesign and then uses the Base cloning route for further content.

The design model and speech model can differ

A voice designed with Qwen3 VoiceDesign is not restricted to Qwen3 for later narration.

Sonano’s supplied reference voices, saved designed voices and cloned voices share a library. Qwen3-TTS Base, VoxCPM2, OmniVoice and Chatterbox can use those assets within their supported language ranges. The origin model records where the voice came from; it does not bind the voice to that engine.

The language must match, and the reference audio and text must be available. A Chinese reference voice cannot be used directly with English-only Chatterbox Turbo. Kokoro and Qwen3-TTS CustomVoice use their own native presets rather than consuming this shared reference library directly.

What transfers is the reference material. Different models may reproduce timbre, pronunciation, pauses and emotion differently. Preview again after changing the engine; a saved voice does not guarantee identical output across models.

Three practical combinations

Create a fictional character. Design a Chinese voice with Qwen3 VoiceDesign or VoxCPM2, preview it and save it. Select that voice for speech, then compare Qwen3-TTS Base and VoxCPM2 on the same new passage. Keep the language and script fixed while changing models.

Read new text in your own voice. Record a clear sample, check its transcript and save a cloned voice. Select a reference-based model supporting that language. For an English voice, Chatterbox Turbo can also be included in the preview comparison.

Get a quick reading. Choose a native Kokoro or Qwen3-TTS CustomVoice preset and generate from text. Alternatively, use a supplied reference voice with a shared-library model. Neither route requires you to design or clone a voice first.

Where transcription, alignment and conversion fit

Whisper and Qwen3-ASR turn audio into text. They can help prepare a reference transcript, but you should check it against the recording. Qwen3-ForcedAligner estimates word or character timing from known text and audio for alignment and subtitles; it does not design voices.

Chatterbox VC takes existing speech and a target voice reference to convert the recording’s vocal timbre. Use speech generation for a new written script, and voice conversion when you want to work with speech that has already been recorded.

For your next production, choose where the voice comes from, then select a compatible model, language and script. Continue with the Voice Library documentation or model catalog.