[gemini-tss-2.5-flash] 10 seconds of weird mic-like feedback before line delivery — Reproducible and linked

Gemini API
, ,
1

Generated at temperature=0

This is how the request was generated

const requestConfig: GenerateContentParameters = {
  model: baseModel,
  contents: [{ role: 'user', parts: [{ text: prompt }] }],
  config: {
    temperature: 0, // Hardcoded 0
    responseModalities: ['audio'],
    speechConfig: {
      voiceConfig: {
        prebuiltVoiceConfig: {
          voiceName: voice_key,
        }
      }
    }
  }
};

// console.log('Request config:', JSON.stringify(requestConfig, null, 2));

const result = await vertexClient.models.generateContent(requestConfig);

Parameters:
voice_key: “Iapetus”
prompt: “Speak the text in native-level Korean, at a relaxed pace. Here’s the text: 내가 어제 너무 늦게 잤어.”
model: “gemini-2.5-flash-tts”