Gemini Speech Generation Explained: How to Use It, Japanese Quality, and Pricing (2026)

Introduction

The speech generation feature of the Google Gemini API turns text into natural-sounding AI audio — and its standout trait is that you can direct the delivery itself with natural-language instructions like "cheerfully" or "in a whisper."

This guide walks you through everything you need to get started: using it in Google AI Studio, calling it from the API with code examples, hearing its actual Japanese quality with embedded samples, understanding the pricing model, and seeing how it differs from other engines. All descriptions of external services reflect the state as of this writing (August 2026).

What Is Gemini Speech Generation?

Gemini speech generation is the text-to-speech (TTS) capability offered through the Gemini API and Google AI Studio. As of this writing, it is served by two dedicated TTS models — gemini-2.5-flash-preview-tts and gemini-2.5-pro-preview-tts — both in preview.

Compared with conventional TTS, four traits stand out.

Method 1: Try It in Google AI Studio (No Code)

The easiest entry point is Google AI Studio, which offers a browser-based Generate speech mode. As of this writing you can start for free with a Google account.

1. 1Sign in to Google AI Studio and open the Generate speech mode
2. 2Enter the text to be spoken. Prepend a style instruction like "read this in a bright, upbeat way" to shape the delivery
3. 3Pick a voice (Zephyr, Kore, and so on) and choose single- or multi-speaker mode
4. 4Generate, listen, and download the audio file if you are happy with it

We recommend experimenting with a few voices and style instructions in AI Studio first to see what suits your script. The interface is updated from time to time, so follow the on-screen guidance if details differ.

Method 2: Call It from the API (Python Example)

To integrate speech generation into an app or workflow, use the Gemini API. Below is the basic pattern with Google's official Python SDK (google-genai): specify a dedicated TTS model and set response_modalities to AUDIO.

from google import genai
from google.genai import types
import wave

client = genai.Client(api_key="YOUR_API_KEY")

response = client.models.generate_content(
    model="gemini-2.5-flash-preview-tts",
    contents="明るくポジティブなトーンで読んでください: "
             "こんにちは。本日は新機能の概要をご紹介します。",
    config=types.GenerateContentConfig(
        response_modalities=["AUDIO"],
        speech_config=types.SpeechConfig(
            voice_config=types.VoiceConfig(
                prebuilt_voice_config=types.PrebuiltVoiceConfig(
                    voice_name="Kore"
                )
            )
        ),
    ),
)

data = response.candidates[0].content.parts[0].inline_data.data
with wave.open("output.wav", "wb") as f:
    f.setnchannels(1)
    f.setsampwidth(2)
    f.setframerate(24000)
    f.writeframes(data)

• Style instructions go inside the prompt as plain language (for example, "read this in a cheerful, positive tone")
• For two-speaker dialogue, assign a voice to each speaker via MultiSpeakerVoiceConfig
• If you want to push quality further, gemini-2.5-pro-preview-tts is also available to try (as of this writing)

How Good Is the Japanese? Listen to Real Samples

What matters more than specs is the actual sound. The samples below were generated by Gemini TTS from the same sentence (a short service introduction) with the style instruction "cheerful and positive." No editing or post-processing was applied.

Our editorial impression: what makes Gemini interesting is how much the character shifts between voices even under the identical instruction. Since impressions vary with script and language, try the free sample comparison page — no sign-up required — to judge with your own ears across 7 languages and other engines.

Pricing: Try Free, Then Pay as You Go

As of this writing, you can start experimenting for free in Google AI Studio, and the Gemini API includes a free tier. For production use, the basic model is pay-as-you-go based on the amount of text processed and audio generated.

Because unit prices and free-tier limits are revised over time, we do not list figures here. Always check the current rates on the official Gemini API pricing page.

How It Differs from ElevenLabs, OpenAI, and Google Cloud TTS

Gemini speech generation is distinguished by prompt-directed expressiveness. Here is a rough positioning against the major engines (as of this writing).

Note that within Google itself, Google Cloud Text-to-Speech (Neural2, Chirp3-HD, etc.) and Gemini TTS are separate product lines. For a detailed cross-engine comparison of voice quality, stability, and Japanese naturalness, see our AI voice quality comparison, and to hear the engines side by side, our four-engine listening comparison.

Using Gemini Voices to Narrate Slide Videos

A practical, high-value use of Gemini voices is narrating presentation decks and training slides. Building it yourself, though, means scripting, generating audio per slide, and assembling everything in a video editor.

SpeechSlide AI automates that entire flow. It ships with four AI voice engines — ElevenLabs, Google, Gemini, and OpenAI — so you can simply select Gemini in your project's audio settings and export an MP4 video whose script, generated from your slides, is narrated by a Gemini voice.

Frequently Asked Questions About Gemini Speech Generation

Q. What is the difference between gemini-2.5-flash-preview-tts and gemini-2.5-pro-preview-tts?

A. As of this writing, flash is positioned for speed and cost efficiency while pro leans toward higher quality. Both are preview-stage TTS models, so try each for your use case and confirm the current positioning in the official documentation.

Q. Can audio generated with Gemini be used commercially?

A. Usage rights for generated audio are governed by Google's terms of service, which can change. Before using it in a commercial project, always verify the current scope and conditions in the official terms.

Q. Can I convert a long narration script in one go?

A. There is a limit to how much text a single request can handle (the exact limit changes with model and API updates — see the official docs). For presentations and training content, generating audio per slide or per paragraph is more practical anyway, making tweaks and re-takes far easier.

Q. How fine-grained is the style control?

A. Beyond overall direction like "cheerful" or "serious tone," it responds to instructions about pace and emotional emphasis to a reasonable degree. The effect varies by voice and script, so compare several voices under the same instruction before choosing.

For how to write narration scripts that sound natural, see our guide to creating AI narration.