Choose Flash TTS or Flash-Lite TTS
The model documentation associates Flash TTS with more demanding voice production: longer narration, multiple speakers and difficult pronunciation. Flash-Lite targets fast, high-volume processing such as read-aloud features or frequently generated short announcements. Both use the same basic API structure. For a business, that means evaluating more than which voice sounds nicer in a first sample. Test text length, repeated runs, required languages and generation time. Internal training material has different requirements from an interactive application. Keep the selected model and voice in a documented configuration so the choice can be reproduced and reviewed.
What the release actually makes available
The API changelog dates general availability to September 22. The model page also contains a conflicting last-model-update field, so the dated changelog is used here for the release date. The API documentation lists German among the supported languages for Flash TTS. Language support does not guarantee perfect pronunciation of every place name or technical term. Nor does the release establish that a particular Workspace menu already exposes these models. Check the API or AI Studio and the access conditions that apply to the project rather than transferring availability claims between products.
Example: narrate an internal training module
A company already has an approved script explaining its administration software. Divide it into short chapters with clear filenames. Choose a voice, delivery speed and pronunciation for recurring product names. First generate a chapter containing abbreviations, numbers and a longer list. Have a reviewer listen while following the text. Produce the remaining chapters only after that sample passes. Store the script version together with the audio so later changes do not reach only one format. A synthetic voice does not validate the accuracy of the material it reads, however convincing its delivery sounds.
Put delivery instructions in metadata
Gemini 3.8 TTS treats input text as the transcript to be spoken. Directions such as “Say cheerfully:” or written speaker labels can therefore become audible if left in the actual text. The documentation requires structured speech_metadata for sustained delivery style and speaker assignment. Brief vocal events such as a pause or laugh use their own notation. In multi-speaker requests, every turn must carry the correct assignment. Review existing templates for these differences before upgrading an older integration. A successful API response does not establish that the generated audio contains the intended spoken content.
Handle custom voices and output formats deliberately
Voice Design creates a designed voice; Voice Replication reproduces an existing voice and, according to the release, includes consent verification. Keep those operations distinct within the project. Record which voice may be used for which purpose and who authorised it. Single-response requests also change their default format: the new generation returns WAV rather than the earlier raw PCM format. Applications that previously added their own WAV header need to adjust that processing. Test playback, duration and import into the destination application. Wrapping already packaged audio again can produce unusable files.
Accept audio using a fixed review script
A useful review script contains common technical terms, dates, telephone numbers, amounts and a sentence with a negation. Listen to each variant fully and compare its meaning with the text. Also measure the time needed to correct and regenerate it. For longer content, consistent volume and voice across chapters matter. Keep the written version available as part of accessible business communication. The AI System Check can establish whether a simple read-aloud feature is sufficient or an API production workflow with version control is worthwhile. The result should fit the publication process you already operate.
Questions about voices, German and migration
Is TTS a talking chatbot? No: these models generate audio from supplied text; conversation logic requires other components or a Live model. Can the output speak German? German is documented for Flash TTS, but actual pronunciation still needs review. Is changing the model ID sufficient for migration? No: review text instructions, speaker metadata and audio format. May a familiar person’s voice simply be copied? Technical capability does not grant permission; use approved voices and the prescribed consent process. How do voice and content differ? The voice shapes delivery, while editorial approval still applies to the script.
Keep it verifiable
Primary sources
- Google: Gemini API release notesSource checked:
- Google: Gemini 3.8 Flash TTS and migrationSource checked:
- Google: AI Studio and Gemini API available regionsSource checked:
