Google: Gemini 3.1 Flash TTS Preview

Google Gemini 3.1 Flash TTS Preview is the text-to-speech edition of the Gemini 3.1 Flash Audio family. It accepts text and produces spoken audio with prompt-directed expression. Google describes this generation as an improvement in speech naturalness, controllability and multilingual performance, with native multi-speaker dialogue and more than 70 languages. A practical use is a short narrated lesson, a character exchange or a spoken passage whose delivery matters as much as its wording.

Prepare a finished script and a performance brief. This generation uses natural-language instructions within the text input to guide vocal style, pace and delivery. Google’s release introduces scene direction to establish a situation, speaker Audio Profiles to define characters, and Director’s Notes to guide pace, tone and accent. Keep those parts identifiable when preparing the prompt so the intended words and the performance instructions can be reviewed independently. An economical brief is easier to revise than several competing descriptions of the same voice.

For example, an original rehearsal brief can describe a calm narrator explaining a museum exhibit, with measured phrasing and a warmer tone on the final sentence. The accompanying script should contain the exhibit’s actual names and the precise wording that needs to be voiced. For a dialogue, give the characters distinct profiles and a clear sequence of turns. Use the scene context to explain whether one speaker is reassuring the other, asking a genuine question or responding with restrained surprise. The generated artifact to assess is the audible performance of that script.

Inline audio tags add local control inside the text, allowing a speaker’s expression to change within a sentence rather than relying entirely on the opening direction. Google describes tags as controlling vocal style, pace and delivery. Use them where the performance needs a particular shift and keep the rest of the wording natural to speak. The useful distinction in this edition is between the high-level character or scene brief and the local change in delivery. A short neutral passage can establish the baseline voice before the script adds more emotional direction.

For multilingual work, supply the final-language script and state the desired accent and speaking style in the brief. Google’s announcement documents support across more than 70 languages. As practical editorial preparation, give recurring names one consistent spelling, expand ambiguous abbreviations when one reading is required and check numbers in the spoken result. A language label alone does not settle how a brand name or borrowed term should sound. Compare the passages containing those terms with the intended reading before accepting a long narration.

The exact Gemini API model page lists an input limit of 8,192 tokens and an output limit of 16,384 tokens for this preview model. DeepMind’s native TTS model card separately lists text input up to 16K tokens and audio output of 32K tokens; those model-card figures are distinct from the Gemini API serving limits. Count the performance directions as part of the prepared input as well as the script. The output allowance is expressed in audio tokens, so it does not guarantee a particular recording duration. Organize a long lesson or story into coherent sections, retain the same character brief and listen to the joins. For scripted dialogue, check attribution and turn order in addition to the spoken words and emotional delivery.

Google’s migration documentation identifies the 3.1 default output as headerless raw PCM with the audio/l16 type. The audio bytes need the appropriate playback or container information when preparing an editable recording; adding a WAV filename alone does not create a WAV header. Check the actual returned format before importing the result. Google also states that audio generated by Gemini 3.1 Flash TTS carries an imperceptible SynthID watermark. The watermark concerns identification of generated audio and does not replace a transcript, speaker labels or a listening review.

The 3.1 preparation style differs materially from 3.8. Gemini 3.8 moves sustained directions and speaker identity into structured speech metadata, uses angle-bracket point-event tags and changes completed-request defaults to WAV. Preserve a 3.1 prompt as a 3.1 prompt rather than mixing in that newer request structure. Google’s supported-model table lists single- and multi-speaker synthesis for 3.1, while its designed-voice and voice-replication controls belong to the 3.8 entries. The separate Flash Live edition handles interactive audio input; this TTS edition serves prepared text-to-audio production.

Review the recording against the written script for omitted or repeated words, names, timing of turns and the intended emotional changes. If one sentence needs a different emphasis, revise that local direction and inspect the affected transition. This workflow makes the performance brief useful: it records what should remain stable about the speaker and what should change in that passage, while the script preserves the content the audience needs to hear.