Gemini 3.8 Flash TTS adds prompt-directed speech

Google has added Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS to the Gemini 3.8 TTS docs. The audio-only models are available through the Gemini API and Google AI Studio’s speech playground. Google describes them as its most expressive speech-generation models, with Flash aimed at richer performances and Flash-Lite designed for lower-cost, high-volume output.

The release follows 3.8 Flash’s launch, Google’s third Flash release in six weeks. Developers can use the text model to draft a script, then pass that script to a TTS model for narration within the Gemini API.

Prompt the performance

Gemini 3.1 Flash TTS established the control scheme that the 3.8 models extend. Its more than 200 audio tags let applications direct emotion, pacing, volume, and delivery across more than 70 languages. Google says the new generation retains those controls while improving expressivity and adding the lower-cost Lite tier.

  • Inline tags such as [whispers], [excited], [short pause], and [slow] can change delivery within a sentence.
  • Google AI Studio provides a director-style workflow for defining character Audio Profiles and scene context.
  • Two-speaker generation supports a separate voice and style for each speaker.
  • Audio output uses 24 kHz, 16-bit mono PCM.
  • SynthID embeds a watermark in the output to help identify AI-generated audio.

Wire the two-step pipeline

Applications typically generate or retrieve a script first, add speaker labels and delivery tags, and then call the TTS endpoint. The speech guide shows a two-speaker request using the Python SDK:

from google import genai

client = genai.Client()

tts = client.interactions.create(
    model="gemini-3.8-flash-tts",
    input="""Anya: [excited] Welcome back to the show!
Liam: [warm] Today, we are looking at speech generation.""",
    response_format={"type": "audio"},
    generation_config={
        "speech_config": [
            {"speaker": "Anya", "voice": "Kore"},
            {"speaker": "Liam", "voice": "Puck"},
        ]
    },
)

The response contains generated audio rather than text. Clients receiving raw PCM may need to add a WAV container before sending the file to players or editing software that expects a .wav file.

Match the model to the workload

Gemini 3.1 Flash TTS provides a historical pricing reference: $1 per million input text tokens and $20 per million output audio tokens, with audio billed at 25 tokens per second. At that conversion rate, one minute uses 1,500 output tokens and costs $0.03 for output, plus the smaller input-text charge. The 3.1 offering also included a free tier and a 50% batch discount. Google positions Flash-Lite as the cheaper 3.8 option, although production budgets should use the current 3.8 rates rather than the earlier model’s pricing.

Design around the limits

Voice direction becomes code

Prompt-level direction is the central change for developers. Traditional TTS engines often expose a small set of SSML controls and global settings. Gemini Flash TTS accepts persona descriptions, scene context, speaker assignments, and inline stage directions, giving applications finer control over each take through text.

Pairing the speech models with Gemini 3.8 Flash creates a coherent script-to-audio pipeline. Teams can generate copy, revise its structure, adjust delivery tags, and render the result through one API family, reducing the handoffs required to iterate on scripted narration.