Google: Gemini 3.8 Flash Lite TTS

Google Gemini 3.8 Flash-Lite TTS is the throughput-focused member of Google’s 3.8 text-to-speech family. It takes a written script and produces spoken audio. Google positions it for read-aloud features, everyday single-speaker narration, high-volume audio production and the speech-output stage of voice-agent systems. A useful deliverable is an article read in a consistent voice, a spoken product explanation or a set of short instructional recordings. The script should already contain the words that the listener needs to hear.

The 3.8 family separates the transcript from performance directions. Keep the spoken text verbatim and place sustained emotion, pace, volume or delivery style in speech metadata. A warm introduction, an urgent instruction and a quiet closing can each have a short style description attached to their own turn. Text such as “Say cheerfully” inside the transcript can itself be spoken. Speaker names belong in the speaker metadata rather than as labels in the audible script. This preparation is particularly important when reusing prompts written for Gemini 3.1 Flash TTS.

Use inline angle-bracket tags for brief events at a specific point: a breath, laugh, cough, sigh or pause. For example, the transcript “The next step is simple. <short pause> Choose your preferred language.” keeps the intended pause next to the words it affects. Put a sustained whispered delivery in the style description instead of repeatedly adding event tags. Google recommends English tag names even for scripts in other languages. These controls direct the voice; an applause track or a door slam calls for separate sound production rather than a vocal-event tag.

Google lists 101 supported languages for this edition, including English, Russian, Chinese, Japanese, Korean, French, German, Spanish, Indonesian and Vietnamese. Write the finished script in the intended language, including the desired form of names, abbreviations and numbers. As editorial preparation, spell out an ambiguous abbreviation or the intended reading of a number when the listener must hear one particular form. Listen to an initial passage containing the difficult names and transitions before treating a long recording as finished. A language count does not make all regional pronunciations interchangeable.

The manufacturer supports prebuilt voices, an extended voice library, designed voice personas and voice replication. Select the voice identity separately from temporary acting directions. Reuse that identity and a compact delivery description across sections when the narration should remain consistent. In the documented two-speaker configuration, give each turn an explicit speaker that matches one of the two configured prebuilt voices. Google’s guide handles multi-character scenes with designed or replicated voices by generating the individual turns. This distinction matters when planning a dialogue rather than a single narrator.

For ordinary completed requests, Google’s 3.8 guide specifies WAV audio with a RIFF header, containing 24 kHz mono, 16-bit signed little-endian PCM. Its streaming configuration instead returns headerless raw PCM chunks by default; the guide also documents raw linear PCM, mu-law and A-law encoding choices. Match playback or editing settings to the actual output format. Renaming headerless PCM to a WAV extension does not add the container information. When assembling separately generated turns, keep the audio format consistent and treat each complete WAV as a file with its own header.

The exact model page lists Gemini API limits of 8,192 input tokens and 16,384 output tokens. Audio tokens do not express a fixed number of spoken seconds. Prepare long material in meaningful sections, retaining the same voice and style brief and listening at the joins. These are the manufacturer’s documented serving limits for this model; they describe a different quantity from a generic catalogue context field. Text is the input and audio is the output, so script drafting, translation and transcription are separate preparation or follow-up tasks.

Flash-Lite shares the 3.8 request structure with Gemini 3.8 Flash TTS. The practical choice concerns the performance required: Flash-Lite is aimed at rapid, repeated production, while Flash TTS is the creative tier for difficult pronunciation, nuanced acting and broader dialect coverage. A routine help article and a dramatic two-character scene place different demands on the voice. Check the actual recording for omitted or repeated words, pronunciation, pacing and audible changes between segments before using it in the final listening experience.