voice design gives a text to speech model a written brief for a speaker. you describe audible traits such as accent, pitch, and pace, then the model uses them when it speaks the script. the voice comes from those directions, not from a recording of a particular person.
how does voice design work?
the model receives the script and a separate description of the speaker. it uses the description to guide the sound of the generated speech, drawing on associations between voice traits and audio learned during training. keeping the script fixed while changing the description lets you compare how each brief affects delivery.
in rumik silk's mulberry 1.6, a tts request carries text and a description. the first is the line to speak. the second tells the voice model how it should sound, for example, "an indian voice, warm timbre, conversational pacing."
descriptions work within the range the model learned during training. rumik trained mulberry 1.5 on a curated set of traits and related phrasings, so "low pitch" and "deep voice" can lead toward a similar result. wording alone cannot add a voice trait the model never learned.
write the brief as one sentence with a few traits you can hear. "warm" says something about timbre, "conversational" says something about pace. "a friendly, professional voice" doesn't give a lot for the model to work with.
compare voice descriptions on the same sentence
keep the line fixed: "your lesson on the monsoon starts in ten minutes." then render it with two descriptions:
"an indian voice, warm timbre, conversational pacing, like an e-learning instructor."
"an indian voice, deep timbre, slow pacing, like a corporate training narrator."
compare how the descriptions affect pace and timbre. then try a longer line with a name in it. a short sample may hide a voice that flattens emphasis across longer sentences.
voice design vs preset voices and cloning
the difference begins with what you supply to the tts model.
approach | main input | what to expect |
|---|---|---|
voice design | a written description of speaker traits | a generated voice shaped by traits the model has learned. |
preset voice | a named speaker or voice id | a fixed starting identity that can be reused across requests. |
a reference recording of a specific person | speech intended to resemble that person's voice. |
in silk, speaker pins a preset while description supplies the voice direction. a mulberry model request can use both. the preset gives you a named voice to return to and the description still tells it how to deliver the line.
what to test before choosing a voice
listen for words that lose emphasis or sentences that all land with the same rhythm.
a voice agent needs a different test. play a greeting, a clarification, and a longer answer with the same description on every turn. if the agent moves between languages, listen for a sudden change of speaker at the language boundary. a preset speaker can help keep the identity steady across requests.
the brief is for the speaker. keep it stable while testing different lines. if one name sounds wrong, fix the text or pronunciation rather than rewriting the entire voice description.