Gemini 3.8 Flash TTS (Google)

Overview

Gemini 3.8 Flash TTS is Google’s flagship text-to-speech model for expressive, studio-style voice output. You give it a verbatim script, and it generates spoken audio that follows your delivery direction. It supports inline vocal events like laughs, breaths, and pauses inside the script. That makes it useful when you need narration or dialogue that sounds acted, not flat.

This model is built for long-form stability. It’s designed to keep a voice consistent across longer reads and multi-turn dialogue. The output is a complete speech recording you can use in podcasts, audiobooks, games, and voice experiences.

What you can build

• Audiobook narration with consistent character voice and pacing control
• Podcast segments with host-and-guest dialogue in a single render
• Game NPC lines with acting cues, reactions, and controlled timing
• Voice agent responses that need a specific tone (calm, firm, urgent)
• Accessibility read-aloud for articles, lessons, and short scripts
• Character prototypes for creative teams to review before recording

Inputs

• The exact script you want spoken, provided as plain text. The model treats this text as a verbatim transcript, so write it as you want it read.
• Optional inline vocal tags placed inside the script using angle brackets. Use them for point-in-time sounds and pauses, such as a short pause, a sigh, or laughter.
• Optional backchannel phrases written inside pipe characters. Use this when you want quick interjections that shouldn’t take the full emphasis of a spoken sentence.
• A choice of 1 voice for single-speaker audio, or 2 voices for dialogue. When you select 2 voices, structure the script as a back-and-forth and label each line with the voice name followed by a colon.
• An optional style description, written in natural language. Put sustained direction here, like emotion, pace, energy, or accent. Keep stage directions out of the transcript when you don’t want them spoken.

Outputs

• A generated audio file in WAV format.
• The WAV is a complete, ready-to-save clip that contains the spoken script, including your pauses and vocalizations.
• Audio is typically produced as 16-bit PCM at 24 kHz in mono. This works well for editing, mixing, and post-processing.

Limitations

• This is text-to-speech only. It accepts text input and returns audio output.
• Multi-speaker generation supports exactly 2 speakers per clip. If your script has 3 or more characters, you must generate per character and edit the clips together.
• The underlying model has an 8,192-token input limit. Some hosted integrations also cap input by characters. Split long scripts into smaller sections.
• Language is auto-detected from the script. Heavy code-switching and inconsistent spelling can reduce pronunciation quality.
• Unclear formatting can hurt results. Missing speaker labels, messy line breaks, or long run-on sentences can cause wrong pacing or unintended emphasis.
• If you embed stage directions directly in the transcript, the model can read them aloud. Put delivery direction in the style field instead.
• Output sample rate is fixed. You can’t request a custom sample rate.

Safety & compliance

• Generated audio includes an inaudible SynthID watermark that marks it as AI-generated.
• Don’t use the model to impersonate a real person without clear consent and rights.
• If you use any voice replication features outside this Wiro page, follow the product’s consent checks and regional rules.
• Avoid feeding copyrighted scripts you don’t have rights to. Safety systems may block generation or return incomplete audio when content resembles copyrighted material.