Gemini 3.8 TTS: Flash vs Flash-Lite for voiceovers

This is script-to-audio generation, not live voice chat

The release covers two Gemini API models and a Voices endpoint for browsing and creating voices. Google's documentation also points to Google AI Studio for trying them. TTS means text-to-speech: it reads supplied words, whereas the Live API handles interactive conversations. This announcement does not establish a new voice feature in the consumer Gemini app or Chat AI.

Sources: Google AI for Developers, Google AI for Developers

Choose by the recording you need, then listen to a sample

Google describes different target workloads for the two models. Treat those descriptions as a starting point for evaluation, not a quality ranking.

• Flash TTS: Google's suggested uses include audiobooks, expressive narration, multi-speaker dialogue, and difficult pronunciation or regional dialects. The model ID is gemini-3.8-flash-tts.
• Flash-Lite TTS: Google's suggested uses include routine read-aloud, single-speaker generation, and higher-volume speech production. Its model ID is gemini-3.8-flash-lite-tts.
• The two 3.8 models share a request format, so developers can compare them using the same script and configuration. Moving from the older 3.1 preview still requires migration checks.

Sources: Google AI for Developers, Google AI for Developers

Designing a voice and replicating a person are different choices

Voice design creates a reusable persona from a description, without requiring a recording of a real speaker. Voice replication instead needs two real recordings from the same adult: a reference sample and the required spoken consent statement. Google documents both workflows in AI Studio and the API.

Sources: Google AI for Developers, Google AI for Developers

Keep stage directions out of the spoken script

Google warns that 3.8 TTS can read ordinary inline directions aloud. Put sustained delivery instructions and speaker labels in speech_metadata, separate from the words to speak. Momentary pauses and vocal events use the documented inline tags.

• For dialogue, single-request multi-speaker generation supports up to two speakers using prebuilt voices. Custom-voice dialogue requires generating each speaker's turn separately.
• For integrations, non-streaming requests now return WAV audio by default; streaming returns raw PCM chunks by default. An older integration that adds its own WAV header needs review.

Sources: Google AI for Developers, Google AI for Developers

Try this short script check before producing a longer recording

This is an editorial exercise, not captured output or a live-app test. Use the same prebuilt voice and settings for both models, then compare what you hear with the written reference.

1. Generate the sample without delivery instructions first. Check that the time, quantity, and two room numbers remain distinct and correct.
2. Try one short delivery change through the supported style control. Check that it changes the delivery without adding spoken instructions or dropping words.
3. Before a real project, repeat with a short, non-sensitive excerpt containing your own terminology. Listen through the entire result, including speaker changes and the ending.

Our community workshop starts at nine thirty in the morning. Bring two notebooks and a blue pen. The first session is in room fourteen; the second is in room forty.

Sources: Google AI for Developers

Use Chat AI for the writing stage, without assuming TTS access

Chat AI's writing guide covers shaping a draft, checking unsupported claims, and making a final human edit. That is a useful preparation step for a voiceover script, but writing assistance and audio generation are separate capabilities. Verify the exact voice-generation tool before choosing an app or subscription.

• Prepare and review a script with the writing guide
• Compare multi-model apps by the tools you need

Sources: Chat AI