Google: Gemini 3.8 Flash TTS

Google Gemini 3.8 Flash TTS is the creative voice-performance tier of Google’s 3.8 text-to-speech family. Its emphasis is acoustic fidelity, emotional range, regional accents and continuity of a character’s voice across extended dialogue. Google highlights audiobooks, studio narration, difficult pronunciation and scenes with substantial vocal acting. The output is spoken audio from a prepared text script, suitable for reviewing a narrator’s performance or arranging a scripted conversation alongside other production elements.

Begin with the exact words to be heard. The 3.8 transcript is treated verbatim, while the speaker and sustained performance belong in separate speech metadata. Give each turn a clear speaker identity and, where useful, a compact style direction such as restrained concern or cheerful curiosity. Split a line into separate turns when the emotion or pacing changes substantially. This makes the brief easier to inspect: the transcript contains the dialogue, and the attached direction explains how that part should be performed. Ordinary written instructions placed in the transcript may be read aloud.

Google’s 3.8 voice system includes prebuilt voices, an extended voice library, designed personas and replicated voice identities. Choose the underlying persona before applying temporary acting changes. A stable regional accent or character identity belongs to the chosen voice; the style field controls the current performance. The maker advises reusing that voice with brief directions rather than repeating long Audio Profile or Director’s Notes blocks. Its documented single-request dialogue configuration combines up to two prebuilt speakers. For designed or replicated characters, the guide directs individual turn generation when assembling a multi-character exchange.

Place a brief human vocalization or pause at the relevant point with an angle-bracket tag. An original script example is “I thought you had left. <short pause> <sigh> I kept the light on.” Attach a subdued delivery to the whole turn rather than putting that instruction among the spoken words. Punctuation and capitalization can help phrase a hesitation or emphasize a word. Google also documents pipe-delimited listener reactions for backchannels and overlapping dialogue, and identifies Flash TTS as the preferred 3.8 edition for heavier overlap. Keep such reactions short enough that the listener can still understand the main line.

The manufacturer lists 130 supported languages for Flash TTS and emphasizes regional and minority dialect coverage. A regional performance brief is useful for a localized audiobook character or a narration whose accent matters to the audience. Prepare the script in its final language, with consistent spellings for recurring names and an intentional reading of abbreviations and numbers. As an editorial step, start the listening review with the passages that contain difficult names, dialect changes or emotional transitions. The desired accent and pacing should be assessed in the generated voice rather than assumed from the language label alone.

The 3.8 output guide distinguishes complete files from streamed audio. The completed-request default is a WAV file with a RIFF header: 24 kHz, mono, signed 16-bit little-endian PCM. The streaming default is headerless raw linear PCM with the same sample characteristics. Raw linear PCM, mu-law and A-law are also documented encoding choices. Use the actual encoding information when preparing playback or importing the result into an editor. For a sequence of separate dialogue files, keep their formats consistent and inspect the joins for sudden changes in level, pace or room tone.

Google’s exact model page specifies Gemini API serving limits of 8,192 input tokens and 16,384 output tokens. These are token limits, not a fixed audiobook duration or an exact number of seconds of speech. Divide long scripts at meaningful boundaries and retain the chosen voice identity across sections. In a chapter-length production, listen for continuity at paragraph and scene changes, as well as for word accuracy within each segment. A good practical review records which lines need pronunciation or acting revisions instead of rewriting the entire voice brief after every take.

Flash TTS and Flash-Lite TTS share the 3.8 preparation structure, so the edition choice can follow the work’s demands. Flash TTS emphasizes acting, complex dialogue, difficult pronunciations and broader dialect coverage; Flash-Lite emphasizes throughput for routine spoken content. The expressive controls concern human speech and vocal events. Background effects, a musical score and a finished production mix can be planned as separate assets. Script writing, translation and recognition of recorded speech remain distinct stages around this text-to-audio model.