Voice library, casting and delivery

Save a reusable voice

Give a reference a useful name, language and transcript. A clean recording with one speaker is easier to use than a noisy or mixed clip. Voice design can create a synthetic reference from a description. Audition the saved voice with the model you intend to use.

Reference language and accent

A multilingual cloning model can carry over the accent of its reference recording. Start with a reference in the language you intend to generate and check a short passage. The reference transcript must match the words in that recording.

For a synthetic reference, Qwen VoiceDesign supports ten documented languages, including French, German and Japanese. Generate the reference first, then use Qwen Base or another compatible cloning model. The cloning model still needs to support the target language. A pre-built voice can also be a useful starting point; the examples include both types.

Multiple voices

Add roles to a project's cast and assign passages or subtitle speakers. Check unresolved assignments before generation. Speaker diarization identifies speakers in a recording; it does not choose their output voices for you. Cast speech is generated in separate requests and assembled afterward.

Directions and emotions

Instruction support is model-specific. Some models use a description or per-passage instruction, others use inline tags, an emotion setting or provider voice parameters. Put directions in the supported control rather than ordinary spoken text. Unsupported tags may be read aloud.

Start with a simple direction such as a steady pace and a short pause. Listen for intelligibility before adding more detailed emotion or style instructions.

Pronunciations

Save reviewed pronunciations for names and recurring terms. Match them to the intended language and voice. Keep the normal written form in subtitles and reading text.

A reusable synthetic narrator

1. Choose a voice-design model that supports the intended language, or start with a suitable pre-built voice.
2. Generate a reference passage and listen to the actual recording. Check the language, accent and delivery before reusing it.
3. Keep one speaker and remove unwanted silence or noise. If you trim the recording, update the transcript to match the retained words.
4. Save the reference with a clear name and language, then register it with a compatible speech service where required.
5. Audition the target narration model with that saved reference. A successful design preview does not establish that every cloning engine can use the result.

Check the cast on a small scene

Use a short exchange between two roles. Confirm that every passage has the intended assignment and that the voices remain distinguishable in context. If the project uses one narrator, check that old per-speaker choices are not the active casting mode.

Give delivery instructions through a control the selected model supports. Test a simple direction before adding several emotions or vocal events. Keep an ordinary sentence as the spoken text; a model may pronounce an unsupported tag literally.

See multiple voices and delivery instructions.