Google DeepMind just released a text-to-speech model that treats AI narration like film directing, not robotic text conversion. On April 15, 2026, the company launched Gemini 3.1 Flash TTS, a voice generation system that accepts natural language instructions embedded directly into text prompts. Instead of simply converting words to audio, developers can now write emotional cues, pacing commands, accent specifications, and scene context that shape how the AI actually speaks. The model scored an Elo rating of 1,211 on the Artificial Analysis TTS leaderboard, placing it second globally and in the most attractive performance-to-cost quadrant for production systems .

How Does Directing an AI Voice Actually Work?

The core innovation in Gemini 3.1 Flash TTS is its audio tag system. Using square bracket syntax, developers embed directional commands directly into the text they send to the model. These tags control delivery at a granular level, allowing precise control over how each phrase sounds. The system supports over 200 audio tags organized across key categories including pacing, emotional expression, and pause length .

For example, a developer could write: "[excited] This is the biggest launch of the quarter!" or "[whispers] Don't tell anyone I said this." The model interprets these instructions and adjusts the vocal performance accordingly. This represents a fundamental shift away from traditional text-to-speech APIs that accept raw text and output standardized speech, toward what Google frames as "AI vocal performance" .

Beyond inline tags, Google AI Studio provides a director's chair interface with three additional control layers. Scene direction lets developers write full narrative context, defining the studio environment and dialogue stakes. Speaker-level controls assign specific voices, accents, and personality profiles to different characters. Format templates pre-configure output as podcast conversations, audiobook narration, news broadcasts, language tutoring, or voice assistant styles .

What Languages and Accents Does the Model Support?

Gemini 3.1 Flash TTS supports 70 languages, with 24 designated as high-quality evaluated languages. These include Japanese, Hindi, Arabic, German, French, Spanish, and Portuguese, covering the majority of global digital commerce .

The accent system represents a notable technical departure from previous Google TTS products. Rather than tying accent to language settings, Gemini 3.1 Flash TTS treats accent as a style prompt controlled through the tag system. This means developers can write English text and instruct the model to deliver it with a Transatlantic accent, Southern US cadence, or Brixton (South London) tone, all within a single API call. Available English accents include American Valley, American Southern, British RP, British Brixton, and Transatlantic .

How to Build Multi-Character Dialogue With AI Voices

• Native Multi-Speaker Support: Gemini 3.1 Flash TTS handles multi-character dialogue natively within a single prompt, eliminating the need for separate API calls for each speaker that previous TTS systems required.
• Character Consistency Across Turns: The model maintains in-character consistency for each speaker, allowing characters to react to each other naturally rather than responding to isolated text snippets.
• Reduced Engineering Overhead: Developers no longer need workarounds to manage response latency, call overhead, or lack of shared conversational context between characters in dialogue-heavy applications.

These capabilities unlock concrete use cases: podcast generation with multiple hosts, audiobook narration with distinct character voices, interactive voice assistants with separate agent personalities, and dramatic scripts for gaming or entertainment. For any application requiring two or more distinct voices, native multi-speaker support eliminates significant engineering complexity .

How Does This Compare to Existing AI Voice Tools?

On the Artificial Analysis TTS leaderboard, a benchmark built on thousands of blind human preference votes, Gemini 3.1 Flash TTS achieved an Elo rating of 1,211 as of April 15, 2026. ElevenLabs holds the top position, while every other major TTS provider, including OpenAI and Amazon Polly, ranks below Gemini 3.1 Flash TTS on this human-preference metric .

More significantly, Artificial Analysis placed Gemini 3.1 Flash TTS in its most attractive quadrant, the zone where high-quality speech output meets low cost-per-request. This metric matters more for production systems than raw benchmark rankings alone .

The practical advantage extends beyond raw quality scores. A model with an Elo rating of 1,200 and 200 audio tags provides more directorial control than a higher-rated model without tag support, because developers gain the ability to shape vocal performance precisely. The ranking ceiling may be lower today, but the capability floor is significantly higher .

How Can Developers Access Gemini 3.1 Flash TTS?

Google launched Gemini 3.1 Flash TTS in public preview across four access paths on April 15, 2026. The fastest path for testing is Google AI Studio at aistudio.google.com, where developers can select the Audio Playground and choose gemini-3.1-flash-tts-preview as the model. This requires no code and is ideal for prototyping voice designs before production deployment .

For production integration, developers can access the model via the Gemini API using the model ID gemini-3.1-flash-tts-preview. The google-genai Python SDK provides programmatic access. Unlike standard Gemini API calls that return text, the TTS endpoint returns audio file output only, requiring developers to structure prompts with scene direction at the top and speaker profiles below .

This release positions Google's text-to-speech capabilities as a directorial tool rather than a simple conversion utility. For developers building podcasts, audiobooks, interactive voice applications, or any audio-first experience, the combination of high-quality output, granular control through audio tags, native multi-speaker support, and global language coverage represents a significant shift in what's possible with AI-generated speech.