Gemini TTS Prompting Guide

Put spoken content in the script field. Use the voice description for the way it should sound, and scene direction for the situation around a conversation. If you put a sentence such as “read this gently” in the spoken script, you risk asking the model to say the instruction aloud. Keeping these jobs separate makes a draft easier to understand and easier to revise when a take needs work.

Start with two or three sentences that represent the final material. Include any name, technical word, or transition that might cause trouble. Listen once for overall tone, then again for individual words. A short test gives you a clear question to answer: does this speaker sound appropriate for this passage? You do not need to configure every control before finding out.

The working method throughout this Gemini TTS prompting guide is simple: write a clear passage, choose a voice, add one useful direction, and listen. Keep the text unchanged while comparing voices. Keep the voice unchanged while comparing styles. This makes it easier to identify the setting that caused an improvement instead of guessing between several simultaneous changes.

A good voice description names a delivery the listener could recognise. “Friendly teacher, clear and patient” is easier to interpret than a long collection of adjectives. For a story, try “quiet narrator, measured pace, curious at the end.” For a tutorial, try “calm explanation, crisp instructions, brief pauses between steps.” These examples describe a performance without changing the words being spoken.

Avoid contradictory directions such as very fast, extremely relaxed, energetic, whispered, and flat all at once. Decide which quality matters most for the scene. If you need a change partway through a long reading, consider making separate segments with distinct direction. That is easier to review than asking one large block of instructions to govern several different scenes.

A voice can sound concerned without sounding theatrical. Give it a reason for the emotion: a reassuring explanation to a worried listener, or a curious question asked during an interview. Brief scene context often communicates more than repeatedly demanding a stronger emotional effect. Keep the feeling appropriate to the text; an upbeat promotional tone can distract from a careful safety explanation or a reflective story.

Pace is also about sentence structure. If a line feels rushed, shorten the sentence or split an idea into two parts before choosing a slower setting. If it feels lifeless, remove unnecessary qualifications and put the main action earlier. A natural recording usually begins with language that a person could comfortably say. This Gemini TTS prompting guide treats script editing as part of voice direction, not an afterthought.

The studio provides <laugh>, <sigh>, <breath>, <short pause>, and <long pause>. Insert a tag at the point where you want that cue in the performance. The buttons beside the script insert at the current cursor or selected text. In dialogue mode, use the buttons belonging to the relevant turn so the cue goes to the intended speaker.

Use <short pause> between a question and its explanation, or before a detail that deserves attention. Use <long pause> for a larger break in thought. These are qualitative cues rather than exact timestamps. If your video needs a precisely measured gap, trim or space the finished audio in your editor after listening to the generated result.

Use <breath> for a breath within a natural reading. Use <sigh> when the character’s reaction calls for it, and <laugh> for a moment of amusement. A tag is a request to the model, not a guarantee of an identical sound on every run. Begin with one cue, listen, and remove it if it distracts from the sentence.

Do not substitute bracketed spellings from an older tutorial and assume they behave the same way. This Gemini TTS prompting guide uses the angle-bracket tags exposed by the current studio. Tags contribute to the submitted script length. Avoid filling a paragraph with cues just because the controls make them easy to insert; each one should serve the spoken moment.

Switch to Two speakers, choose a voice for each role, and assign each line to speaker one or speaker two. Write only the spoken words in each turn. You do not need to prefix the text with “Speaker 1:” because the interface already stores that assignment. A shared scene description can explain where the conversation takes place and how the speakers relate to each other.

Give each participant a consistent role. An interviewer might ask short, curious questions while a guest gives measured answers. A learner might express uncertainty while a teacher explains one idea at a time. A useful contrast comes from their purpose as well as their voices. Avoid making both speakers repeat the same information simply to extend the exchange.

Read the dialogue aloud in order. Every answer should respond to the preceding line, and each pause should help the exchange feel intentional. The application supports two speakers per generation, so a script with a larger cast needs to be adapted or produced in separate requests. Those requests are not automatically merged into a finished episode.

If a name sounds wrong, start by checking the spelling and nearby punctuation. Try a clearer written form in a short test, then listen again. If a sentence sounds too intense, simplify the voice direction or remove a vocal cue. If the recording sounds monotonous, first check whether the script has natural sentence variation and a clear point of emphasis.

If a turn is missing or assigned incorrectly, check the speaker selector on every line and make sure no required turn is empty. Review the displayed character count before submitting. The studio accepts up to 5,000 characters per generation; long passages should be split at meaningful boundaries. Dialogue length includes the submitted turns rather than treating every speaker as a separate allowance.

Keep a simple record of the chosen voice, model, direction, and the exact text used for a successful take. This Gemini TTS prompting guide provides starting examples, but your own reviewed settings are the best reference for the next segment of the same project. Consistency is easier when the inputs are deliberate and the finished audio is checked.