Hi,
We are building a phone voice agent (Uzbek and Russian) on gemini-3.8-live through the Gemini Developer API (AI Studio API key, not Vertex AI). We need the agent to speak with our own brand voice.
What we do:
- google-genai 2.30.1, client.aio.live.connect(model=“gemini-3.8-live”, …)
- speech_config.voice_config.replicated_voice_config with mime_type=“audio/wav” and voice_sample_audio = a 16-second, 24 kHz, 16-bit mono WAV of clean speech.
What happens:
- The session connects with no error and the model answers with audio.
- But the output voice does not change. It sounds like the default prebuilt voice (close to “Kore”).
- A white-noise sample is also accepted without any error, so the reference does not seem to be validated or used.
- A/B test with the same prompt: Kore = 217 Hz median F0, Puck = 103 Hz, our replicated voice = 181–226 Hz, a deliberately lower-pitched reference (163 Hz) still gave 211 Hz. So the reference does not steer the voice.
Questions:
- Is replicated_voice_config for gemini-3.8-live supported on the Gemini Developer API, or only for allowlisted Vertex AI customers?
- If access is needed, how can we request it?
- Can a voice created with POST /v1beta/voices (voice_… id) be used with gemini-3.8-live? The docs list only gemini-3.8-flash-tts and gemini-3.8-flash-lite-tts.
- Is Uzbek (uz-UZ) planned as a consent locale for voice replication?
Thank you!