Google's Gemini 3.1 Flash TTS puts developers in the 'director's chair' with inline audio tags

What the audio tags actually change

The headline feature of Gemini 3.1 Flash TTS is not raw voice quality — it's the control surface. Google describes 'audio tags,' natural-language commands embedded directly into the text input to steer vocal style, pace, and delivery. Instead of picking from a fixed menu of voices and prosody settings, a developer writes the direction inline alongside the words being spoken.

Google frames this as three layers: scene direction that sets an environment and keeps characters 'in-character' across multiple turns; speaker-level specificity through unique Audio Profiles plus Director's Notes for pace, tone, and accent; and inline tags that let a speaker 'pivot from these high-level settings to change expression mid-sentence.' That mid-sentence granularity is the meaningful shift — it moves TTS from clip-level configuration toward performance-level directing.

Once the performance is perfected, these exact parameters can be exported as Gemini API code to ensure consistent, recognizable voices across various projects and platforms.Montana Labs

The export path is the part teams should test first

For applied teams, the interesting workflow claim is the one above: you tune a voice interactively in the Google AI Studio Playground, then export the exact parameters as Gemini API code. This addresses a real friction point — the gap between what sounds right in a playground and what you can reliably reproduce in production. If the export genuinely captures Audio Profiles, Director's Notes, and inline tags as reusable code, it turns voice design into a versionable artifact rather than a set of settings someone remembers.

The rollout is staged and still in preview: developers get it via the Gemini API and Google AI Studio, enterprises via Vertex AI, and Workspace users through Google Vids. The three surfaces suggest Google wants the same model feeding both hands-on developer tooling and packaged consumer-facing products like Vids.

The benchmark and cost positioning

Google cites an Elo score of 1,211 on the Artificial Analysis TTS leaderboard, a benchmark built on thousands of blind human preferences, and notes the model sits in that benchmark's 'most attractive quadrant' for combining high-quality generation with low cost. The 'Flash' branding signals the same intent: this is positioned as the fast, affordable tier rather than a premium quality-at-any-price model. For teams weighing per-character or per-second costs at volume, the low-cost framing matters as much as the expressivity.

Native multi-speaker dialogue and support for 70+ languages round out the pitch. Google explicitly ties the language coverage to localization — bringing style, pacing, and accent control to 'major markets' so a single model can produce expressive speech across regions rather than English-first output with everything else as an afterthought.

Every clip carries a SynthID watermark

Google states that all audio generated by 3.1 Flash TTS is watermarked with SynthID, an imperceptible marker 'interwoven directly into the audio output' to allow reliable detection of AI-generated content. Notably, this is presented as non-optional — a property of every generation, not a toggle.

That default matters precisely because the model's selling point is expressive, natural, multi-speaker human-sounding voices. The more convincing the output, the more the watermark becomes part of the product rather than a compliance footnote. Teams building on this should assume detectability is baked in, and plan for downstream systems — moderation, provenance checks, platform review — that can read it. The implication of this release is that Google is shipping fine-grained voice performance and traceable provenance as a single package, and betting developers will accept the second to get the first.