Gemini 3.8 Voice Replication: POST /v1beta/voices returns 404 "Requested entity was not found." for every input (same audio worked on Sep 26)

Gemini API
1

Since today (2026-10-01), every Voice Replication request I send fails with HTTP 404 NOT_FOUND “Requested entity was not found.” A viewer of my tutorial hit the same message in AI Studio’s Voice Replication UI, so it is not limited to REST. I reproduced it several times, most recently at 14:13 UTC.

Environment

  • Gemini Developer API, REST: POST https://generativelanguage.googleapis.com/v1beta/voices
  • Paid tier (project with billing enabled)
  • Requests sent from South Korea

What fails (all return 404 “Requested entity was not found.” in 0.4–0.7 s)

  • The exact reference (21.1 s, 24 kHz mono PCM16) + Korean consent (6.6 s) pair that successfully created a stored voice on 2026-09-26 (CreateVoice returned 200 in about 7 s back then)
  • store: true and store: false
  • "model": "gemini-3.8-flash-tts", and with voice.model omitted
  • 1 s of silence as both clips (so the request fails before any audio validation)

What still works with the same key

  • GET /v1beta/voices and GET /v1beta/voices/{id} for the voice created on Sep 26: 200
  • Speech synthesis via /v1beta/interactions with that replicated voice: 200
  • Voice Design (type: "prompted", store: true): 200

Other observations

  • With "model": "gemini-3.8-flash-lite-tts" and type: "replicated", the API returns 400: “Model ‘gemini-3.8-flash-lite-tts’ does not support Voice Replication. Omit ‘voice.model’ to use the default voice creation model; the resulting voice can then be used for speech synthesis with any TTS model.” The voice-replication docs still say both Flash and Flash-Lite support replication and show "model": "gemini-3.8-flash-tts" in the request. Omitting voice.model as the message suggests still gives the same 404.
  • A wrong model name or an unknown voice id returns a different, specific 404 message, so this generic one seems to come from the voice creation step itself.

Request (audio omitted)

{
  "store": false,
  "voice": {
    "model": "gemini-3.8-flash-tts",
    "type": "replicated",
    "display_name": "probe",
    "replicated": {
      "source_audio": {"mime_type": "audio/wav", "data": "<base64>"},
      "consent_audio": {"mime_type": "audio/wav", "data": "<base64>"}
    }
  }
}

Response

{"error": {"code": 404, "message": "Requested entity was not found.", "status": "NOT_FOUND"}}

Is creating new replicated voices currently broken or disabled? Is there a workaround?