{ "@context": "https://schema.org", "@id": "https://aggr.aymericbeaumet.com/items/philschmid-de/2026-09-24-create-your-own-voice-for-gemini-3-8-tts/#webpage", "@type": [ "WebPage", "ArchiveComponent" ], "articleBody": "[Gemini 3.8 Flash TTS and Flash-Lite TTS](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-text-to-speech/) are now available, in the Gemini API and [AI Studio](https://aistudio.google.com/generate-speech). Ranking #1 on Hume's Voice Design Benchmark, and on top of [Voice Arena](https://voicearena.com/tts-leaderboard/us-english) in 6 languages.\n\nBiggest 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.\n\n### 1\\. Record 2 clips\n\nSame mic, same room, 24kHz mono. The API compares the two recordings, so don't switch from a headset to the laptop mic between them.\n\nBash\n\n```bash\n# macOS. \":0\" is audio device 0. To find your mic:\n# ffmpeg -f avfoundation -list_devices true -i \"\"\n \n# me.wav: 15-20s of you talking like you normally do. Explain what\n# you're building this week. Don't read, just talk.\nffmpeg -f avfoundation -i \":0\" -ac 1 -ar 24000 -t 20 me.wav\n \n# consent.wav: read this word for word:\n# \"I am the owner of this voice and I consent to Google using this voice to create a synthetic voice model.\"\nffmpeg -f avfoundation -i \":0\" -ac 1 -ar 24000 -t 8 consent.wav\n```\n\nLinux: `-f alsa -i default`. On macOS, the first run may only open the microphone permission dialog. Allow it, then run the command again.\n\nThe consent sentence must be one of the [25 supported versions](https://ai.google.dev/gemini-api/docs/voice-replication#consent-phrases-by-language), 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.\n\n### 2\\. Create the voice, then speak\n\nPython\n\n```python\nimport base64\nfrom google import genai\n \nclient = genai.Client()\n \ndef b64(path):\n return base64.b64encode(open(path, \"rb\").read()).decode()\n \nvoice = client.voices.create(\n store=True, # kept in your project for 1 year, returns voice_...\n voice={\n \"model\": \"gemini-3.8-flash-tts\",\n \"type\": \"replicated\",\n \"display_name\": \"Me\",\n \"replicated\": {\n \"source_audio\": {\"mime_type\": \"audio/wav\", \"data\": b64(\"me.wav\")},\n \"consent_audio\": {\"mime_type\": \"audio/wav\", \"data\": b64(\"consent.wav\")},\n },\n },\n)\nprint(voice.id)\n \ninteraction = client.interactions.create(\n model=\"gemini-3.8-flash-tts\",\n input=[{\n \"type\": \"user_input\",\n \"content\": [{\n \"type\": \"text\",\n \"text\": \"Okay so... I did not record this. \"\n \"Twenty seconds of audio and one consent sentence. That's it.\",\n \"annotations\": [{\"type\": \"speech_metadata\", \"style\": \"casual, a bit amused\"}],\n }],\n }],\n response_format={\"type\": \"audio\"},\n generation_config={\"speech_config\": [{\"voice\": voice.id}]},\n)\n \n# 3.8 returns a real WAV with a RIFF header. No wave-module wrapping.\nopen(\"me_synth.wav\", \"wb\").write(base64.b64decode(interaction.output_audio.data))\n```\n\nTwo 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 `` and ``. The voice ID is reusable: pass it in any later request, or find it again with `client.voices.list(type_=[\"replicated\"])`.\n\nIf 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.\n\n## Or design one from a sentence\n\nNo 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:\n\nPython\n\n```python\nvoice = client.voices.create(\n store=True,\n voice={\n \"model\": \"gemini-3.8-flash-tts\",\n \"type\": \"prompted\",\n \"display_name\": \"Deadpan host\",\n \"gender\": \"male\",\n \"language_code\": \"en-US\",\n \"prompted\": {\"input\": \"A dry, deadpan podcast host in his 30s, low pitch, slight German accent.\"},\n },\n)\nopen(\"preview.wav\", \"wb\").write(base64.b64decode(voice.sample_audio.data))\n```\n\nThen use `voice.id` exactly as above.\n\n## What changed in prompting\n\nIf you're coming from `gemini-3.1-flash-tts-preview`, this will break your prompts, so read the [prompting guide](https://ai.google.dev/gemini-api/docs/speech-generation#prompting-guide) and the [migration notes](https://ai.google.dev/gemini-api/docs/speech-generation#migration). The short version:\n\n- Input text is spoken word for word. Instructions written inside the text get spoken out loud too.\n- Delivery that lasts the whole line (\"whispering\", \"out of breath\") goes in `speech_metadata.style`. Short sounds go inline in angle brackets.\n- Multi-speaker turns need an explicit `speaker` on every turn.\n- Long \"Audio Profile\" prompts now cause voice drift. Design the persona once with the Voices API, then send short or empty `style` strings.\n- 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.\n", "authors": [], "captured_at": "2026-09-27T19:29:15.293681927+00:00", "category": "programming", "content_html": "

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.

\n

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.

\n

1. Record 2 clips

\n

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.

\n

Bash

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

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

\n

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.

\n

2. Create the voice, then speak

\n

Python

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

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"]).

\n

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.

\n

Or design one from a sentence

\n

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:

\n

Python

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

Then use voice.id exactly as above.

\n

What changed in prompting

\n

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:

\n
    \n
  • Input text is spoken word for word. Instructions written inside the text get spoken out loud too.
    • \n
  • Delivery that lasts the whole line ("whispering", "out of breath") goes in speech_metadata.style. Short sounds go inline in angle brackets.
    • \n
  • Multi-speaker turns need an explicit speaker on every turn.
    • \n
  • Long "Audio Profile" prompts now cause voice drift. Design the persona once with the Voices API, then send short or empty style strings.
    • \n
  • 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.
    • \n
\n", "content_markdown": "[Gemini 3.8 Flash TTS and Flash-Lite TTS](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-text-to-speech/) are now available, in the Gemini API and [AI Studio](https://aistudio.google.com/generate-speech). Ranking #1 on Hume's Voice Design Benchmark, and on top of [Voice Arena](https://voicearena.com/tts-leaderboard/us-english) in 6 languages.\n\nBiggest 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.\n\n### 1\\. Record 2 clips\n\nSame mic, same room, 24kHz mono. The API compares the two recordings, so don't switch from a headset to the laptop mic between them.\n\nBash\n\n```bash\n# macOS. \":0\" is audio device 0. To find your mic:\n# ffmpeg -f avfoundation -list_devices true -i \"\"\n \n# me.wav: 15-20s of you talking like you normally do. Explain what\n# you're building this week. Don't read, just talk.\nffmpeg -f avfoundation -i \":0\" -ac 1 -ar 24000 -t 20 me.wav\n \n# consent.wav: read this word for word:\n# \"I am the owner of this voice and I consent to Google using this voice to create a synthetic voice model.\"\nffmpeg -f avfoundation -i \":0\" -ac 1 -ar 24000 -t 8 consent.wav\n```\n\nLinux: `-f alsa -i default`. On macOS, the first run may only open the microphone permission dialog. Allow it, then run the command again.\n\nThe consent sentence must be one of the [25 supported versions](https://ai.google.dev/gemini-api/docs/voice-replication#consent-phrases-by-language), 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.\n\n### 2\\. Create the voice, then speak\n\nPython\n\n```python\nimport base64\nfrom google import genai\n \nclient = genai.Client()\n \ndef b64(path):\n return base64.b64encode(open(path, \"rb\").read()).decode()\n \nvoice = client.voices.create(\n store=True, # kept in your project for 1 year, returns voice_...\n voice={\n \"model\": \"gemini-3.8-flash-tts\",\n \"type\": \"replicated\",\n \"display_name\": \"Me\",\n \"replicated\": {\n \"source_audio\": {\"mime_type\": \"audio/wav\", \"data\": b64(\"me.wav\")},\n \"consent_audio\": {\"mime_type\": \"audio/wav\", \"data\": b64(\"consent.wav\")},\n },\n },\n)\nprint(voice.id)\n \ninteraction = client.interactions.create(\n model=\"gemini-3.8-flash-tts\",\n input=[{\n \"type\": \"user_input\",\n \"content\": [{\n \"type\": \"text\",\n \"text\": \"Okay so... I did not record this. \"\n \"Twenty seconds of audio and one consent sentence. That's it.\",\n \"annotations\": [{\"type\": \"speech_metadata\", \"style\": \"casual, a bit amused\"}],\n }],\n }],\n response_format={\"type\": \"audio\"},\n generation_config={\"speech_config\": [{\"voice\": voice.id}]},\n)\n \n# 3.8 returns a real WAV with a RIFF header. No wave-module wrapping.\nopen(\"me_synth.wav\", \"wb\").write(base64.b64decode(interaction.output_audio.data))\n```\n\nTwo 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 `` and ``. The voice ID is reusable: pass it in any later request, or find it again with `client.voices.list(type_=[\"replicated\"])`.\n\nIf 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.\n\n## Or design one from a sentence\n\nNo 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:\n\nPython\n\n```python\nvoice = client.voices.create(\n store=True,\n voice={\n \"model\": \"gemini-3.8-flash-tts\",\n \"type\": \"prompted\",\n \"display_name\": \"Deadpan host\",\n \"gender\": \"male\",\n \"language_code\": \"en-US\",\n \"prompted\": {\"input\": \"A dry, deadpan podcast host in his 30s, low pitch, slight German accent.\"},\n },\n)\nopen(\"preview.wav\", \"wb\").write(base64.b64decode(voice.sample_audio.data))\n```\n\nThen use `voice.id` exactly as above.\n\n## What changed in prompting\n\nIf you're coming from `gemini-3.1-flash-tts-preview`, this will break your prompts, so read the [prompting guide](https://ai.google.dev/gemini-api/docs/speech-generation#prompting-guide) and the [migration notes](https://ai.google.dev/gemini-api/docs/speech-generation#migration). The short version:\n\n- Input text is spoken word for word. Instructions written inside the text get spoken out loud too.\n- Delivery that lasts the whole line (\"whispering\", \"out of breath\") goes in `speech_metadata.style`. Short sounds go inline in angle brackets.\n- Multi-speaker turns need an explicit `speaker` on every turn.\n- Long \"Audio Profile\" prompts now cause voice drift. Design the persona once with the Voices API, then send short or empty `style` strings.\n- 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.\n", "dateCreated": "2026-09-27T19:29:15.293681927+00:00", "date_created": "2026-09-24T00:00:00+00:00", "date_modified": "2026-09-24T00:00:00+00:00", "date_published": "2026-09-24T00:00:00+00:00", "description": "Gemini 3.8 Flash TTS lets you replicate your own voice or design one from a sentence, then direct delivery line by line.", "external_url": "https://www.philschmid.de/gemini-3-8-tts", "git": { "commit": "ebb0941c5c06a2ab2377342d607535262a3e969e", "path": "items/philschmid-de/2026/09/2026-09-24-create-your-own-voice-for-gemini-3-8-tts.md", "permalink": "https://github.com/aymericbeaumet/aggr-instance/blob/ebb0941c5c06a2ab2377342d607535262a3e969e/items/philschmid-de/2026/09/2026-09-24-create-your-own-voice-for-gemini-3-8-tts.md" }, "id": "urn:aggr:item:235cd583c3dd7fe71f44a792b22d3d1145ed3332", "isBasedOn": { "@id": "https://www.philschmid.de/gemini-3-8-tts" }, "mainEntity": { "@id": "https://www.philschmid.de/gemini-3-8-tts", "@type": "CreativeWork", "archivedAt": { "@id": "https://aggr.aymericbeaumet.com/items/philschmid-de/2026-09-24-create-your-own-voice-for-gemini-3-8-tts/#webpage" }, "dateModified": "2026-09-24T00:00:00+00:00", "datePublished": "2026-09-24T00:00:00+00:00", "headline": "Create your own voice for Gemini 3.8 TTS", "timeRequired": "PT4M", "url": "https://www.philschmid.de/gemini-3-8-tts", "wordCount": 680 }, "name": "Archived snapshot: Create your own voice for Gemini 3.8 TTS", "preview": null, "reading_minutes": 4, "replicated_at": null, "source": "philschmid-de", "source_name": "philschmid.de - RSS feed", "summary": "Gemini 3.8 Flash TTS lets you replicate your own voice or design one from a sentence, then direct delivery line by line.", "tags": [], "temporalCoverage": "2026-09-27T19:29:15.293681927+00:00", "title": "Create your own voice for Gemini 3.8 TTS", "url": "https://aggr.aymericbeaumet.com/items/philschmid-de/2026-09-24-create-your-own-voice-for-gemini-3-8-tts/", "word_count": 680 }