← Back to feed

Google DeepMind released Gemini 3.1 Flash TTS on April 15 in developer preview, available immediately through the Gemini API and Google AI Studio. Enterprise access goes through Vertex AI; Google Workspace users get it embedded in Google Vids.

The launch is notable for two things: benchmark position and control architecture.

Leaderboard Position

On the Artificial Analysis TTS leaderboard — which aggregates thousands of blind human preference votes — Gemini 3.1 Flash TTS scores an Elo of 1,211. That puts it second overall, above ElevenLabs v3 and behind Inworld 1.5 Max. At $1.00 per million text tokens and $20.00 per million audio tokens, the quality-to-cost ratio is one of the best on the market. Batch mode halves both figures to $0.50/M and $10.00/M respectively, with Google not using batch-mode data for product improvement.

This matters: ElevenLabs has dominated the TTS quality narrative for two years. A Google model beating it on a neutral third-party benchmark at lower pricing is a direct threat to ElevenLabs’ enterprise business.

Audio Tags: The Control Architecture

The most technically significant feature is audio tags. Rather than static voice configuration, developers embed inline modifiers directly into the transcript text: [whispers], [laughs], [sighs], [panicked], [sarcastic], [gasp]. The model interprets these at generation time, adjusting delivery for those lines without affecting the rest of the output.

Combined with an “audio profile” layer — defining who is speaking, their character archetype, and delivery characteristics — and a scene context block, developers get three levels of control:

1. Global: audio profile sets the baseline voice and character
2. Contextual: scene description keeps multi-turn dialogue in character
3. Granular: audio tags override delivery for specific lines

Director’s notes can be exported as Gemini API code, making it repeatable across deployments. The prompting structure is documented and openly explained, not a black box.

Scope and Availability

Gemini 3.1 Flash TTS supports more than 70 languages including Japanese, Hindi, and German, with regional accent variants available. Native multi-speaker dialogue — multiple voices in a single generation pass — is included. All output is watermarked with SynthID for provenance tracking.

A free tier exists but Google uses that data for model improvement. Paid tiers offer the same model without data retention for training.

Context

The voice API market is entering a competitive phase. xAI launched its own Grok STT and TTS APIs the same week. OpenAI’s audio endpoints run inside the broader GPT-5 API stack. Google is now competing directly on the specialist TTS benchmarks that ElevenLabs has used as proof of quality.

The pattern is consistent with what has happened in every other AI capability category: frontier labs commoditise the feature into a line-item on their API bill, and specialised providers need to differentiate on depth of control, latency, or integration rather than raw output quality alone.