Create your own voice for Gemini 3.8 TTS

Gemini 3.8 Flash TTS and Flash-Lite TTS are now available, in the Gemini API and AI Studio. Ranking #1 on Hume's Voice Design Benchmark, and on top of Voice Arena in 6 languages.

Biggest new features is that you can now replicate your own voice or create a new one from a sentence. The guide below shows how to do both.

1. Record 2 clips

Same mic, same room, 24kHz mono. The API compares the two recordings, so don't switch from a headset to the laptop mic between them.

# macOS. ":0" is audio device 0. To find your mic:
# ffmpeg -f avfoundation -list_devices true -i ""
 
# me.wav: 15-20s of you talking like you normally do. Explain what
# you're building this week. Don't read, just talk.
ffmpeg -f avfoundation -i ":0" -ac 1 -ar 24000 -t 20 me.wav
 
# consent.wav: read this word for word:
# "I am the owner of this voice and I consent to Google using this voice to create a synthetic voice model."
ffmpeg -f avfoundation -i ":0" -ac 1 -ar 24000 -t 8 consent.wav

Linux: -f alsa -i default. On macOS, the first run may only open the microphone permission dialog. Allow it, then run the command again.

The consent sentence must be one of the 25 supported versions, read exactly. German, for example: "Ich bin der Eigentümer dieser Stimme und bin damit einverstanden, dass Google diese Stimme zur Erstellung eines synthetischen Stimmmodells verwendet." The reference clip can be in any language. Use the language you want the voice to speak later.

2. Create the voice, then speak

import base64
from google import genai
 
client = genai.Client()
 
def b64(path):
    return base64.b64encode(open(path, "rb").read()).decode()
 
voice = client.voices.create(
    store=True,  # kept in your project for 1 year, returns voice_...
    voice={
        "model": "gemini-3.8-flash-tts",
        "type": "replicated",
        "display_name": "Me",
        "replicated": {
            "source_audio": {"mime_type": "audio/wav", "data": b64("me.wav")},
            "consent_audio": {"mime_type": "audio/wav", "data": b64("consent.wav")},
        },
    },
)
print(voice.id)
 
interaction = client.interactions.create(
    model="gemini-3.8-flash-tts",
    input=[{
        "type": "user_input",
        "content": [{
            "type": "text",
            "text": "Okay so... <short pause> I did not record this. <laugh> "
                    "Twenty seconds of audio and one consent sentence. That's it.",
            "annotations": [{"type": "speech_metadata", "style": "casual, a bit amused"}],
        }],
    }],
    response_format={"type": "audio"},
    generation_config={"speech_config": [{"voice": voice.id}]},
)
 
# 3.8 returns a real WAV with a RIFF header. No wave-module wrapping.
open("me_synth.wav", "wb").write(base64.b64decode(interaction.output_audio.data))

Two things to notice. The text is spoken word for word. The delivery ("casual, a bit amused") goes in speech_metadata.style, and the short sounds go inline as <short pause> and <laugh>. The voice ID is reusable: pass it in any later request, or find it again with client.voices.list(type_=["replicated"]).

If you don't want anything stored server-side, store=False returns an encrypted voicekey_... that you keep yourself. It works in the same speech_config field and expires after 7 days.

Or design one from a sentence

No recordings needed. Same call, different voice dict. You also get a sample_audio preview, so you can listen to the voice before you synthesize anything:

voice = client.voices.create(
    store=True,
    voice={
        "model": "gemini-3.8-flash-tts",
        "type": "prompted",
        "display_name": "Deadpan host",
        "gender": "male",
        "language_code": "en-US",
        "prompted": {"input": "A dry, deadpan podcast host in his 30s, low pitch, slight German accent."},
    },
)
open("preview.wav", "wb").write(base64.b64decode(voice.sample_audio.data))

Then use voice.id exactly as above.

What changed in prompting

If you're coming from gemini-3.1-flash-tts-preview, this will break your prompts, so read the prompting guide and the migration notes. The short version:

• Input text is spoken word for word. Instructions written inside the text get spoken out loud too.
• Delivery that lasts the whole line ("whispering", "out of breath") goes in speech_metadata.style. Short sounds go inline in angle brackets.
• Multi-speaker turns need an explicit speaker on every turn.
• Long "Audio Profile" prompts now cause voice drift. Design the persona once with the Voices API, then send short or empty style strings.
• Non-streaming (unary) responses are real WAV (audio/wav), so remove any code that adds a WAV header. Set response_format to audio/l16 if you still want raw PCM.