benchr
Gemini 3.1 Flash TTS Preview is not a general Gemini endpoint that happens to speak. Its job is narrower and useful: render supplied text as audio while following natural-language direction about delivery. That boundary should shape both the product and the evaluation.
Choose TTS for scripts; choose Live for conversations
Google's speech guide distinguishes generated speech from the Live API. TTS is for exact text recitation: narration, announcements, accessibility audio, podcast segments, and localized scripts. Live is the path for unstructured two-way audio. The model card does not list Live API support, function calling, structured output, grounding, code execution, or thinking for this TTS model.
This distinction prevents an expensive architecture mistake. A voice assistant that must listen, interrupt, call a tool, and continue is not completed by adding a microphone around this endpoint. A publishing workflow that already owns the final script benefits from the separation: text approval happens before synthesis and the audio artifact can be reviewed, versioned, and regenerated.
The speaker map is part of the prompt
Google documents single-speaker and multi-speaker generation with up to two speakers, along with 30 preset voices. In a two-speaker production, use stable speaker names in the transcript and map each name to one voice. Test overlaps, short acknowledgements, long turns, and a speaker returning after several paragraphs. A beautiful first exchange does not prove that the assignment stays consistent through the episode.
Arabic is in the documented language list. That establishes availability, not quality for every dialect, proper name, or code-switch. The acceptance set should include Modern Standard Arabic, the product's actual regional phrasing, Latin-script brands, numerals, dates, abbreviations, and Arabic-English transitions. Human reviewers should mark the precise phrase that failed and whether the problem is pronunciation, stress, pace, or speaker identity.
| Production need | Documented path | Acceptance evidence |
|---|---|---|
| Single narrator | One preset voice with style instructions | Names, numbers, pacing, and long-form continuity pass |
| Two-person script | Two speaker names mapped to two voices | No speaker swaps or identity drift |
| Arabic localization | Arabic is officially supported | Dialect, code-switch, and proper-name review by native listeners |
| Large offline queue | Batch API | Deterministic asset naming, retry, and review workflow |
| Interactive agent | Use Live API instead | Do not force TTS into a conversation loop |
A workload that exposes editing cost
Prepare a locked script set: a short notification, a dense explanatory paragraph, Arabic and English names, dates and identifiers, an emotional direction, a whispered or sighed tag only where editorially appropriate, a two-speaker exchange, and a long segment that must be split. Generate every item with the same voice mapping and configuration.
Score exact-word fidelity, pronunciation, delivery compliance, speaker consistency, join quality between segments, regeneration count, and reviewer time to an approved asset. The best model is not merely the take listeners prefer; it is the workflow that produces an approvable file without repeated script hacks. Keep the original script, prompt, voice names, model ID, segment order, and accepted output together.
Streaming and Batch solve different queues
Google's changelog records streaming generation through the generation API, while the model card lists Batch support. Streaming helps a product begin playback before the entire artifact is ready; Batch helps a controlled offline queue. Test streaming boundaries for clipped starts, awkward pauses, cancellation, and segment joins. Test Batch for idempotent retries, stable file naming, and partial-job recovery.
Neither mode removes editorial review. Split long scripts at semantic boundaries rather than arbitrary token counts, preserve pronunciation guidance across segments, and listen to the joins in sequence. The documented 8,192 input-token and 16,384 output-token limits are ceilings, not recommended segment sizes.
When to avoid or delay it
Avoid this endpoint for a conversational agent, tool-using voice workflow, or any design that requires live turn-taking. Delay a production commitment if a stable generally available endpoint is mandatory, a public price is required for approval, or your Arabic dialect and proper-name set has not passed native review. Do not borrow a rate from a Gemini text model; the checked TTS page does not publish one.
| Field | Official record | Decision effect |
|---|---|---|
| Input / output | Text / audio | Approve the script before synthesis |
| Limits | 8,192 input / 16,384 output tokens | Segment long scripts deliberately |
| Speech options | 30 voices; up to two speakers; Arabic documented | Test exact voice, language, and speaker map |
| Delivery modes | Streaming generation and Batch | Design separate online and offline recovery paths |
| Price | No public rate recorded on checked page | Keep the production budget open |