## Environment
- SDK: `@google/genai` v1.47.0
- API Version: v1alpha (ephemeral tokens)
- Model: `models/gemini-3.5-live-translate-preview`
- Platform: Web (Chrome), WebSocket via SDK `ai.live.connect()`
-–
## Configuration
Ephemeral token created with liveConnectConstraints:
```javascript
{
model: “models/gemini-3.5-live-translate-preview”,
config: {
responseModalities: [“AUDIO”],
translationConfig: {
targetLanguageCode: “es”,
echoTargetLanguage: false
}
}
}
```
Client connect config (same values):
```javascript
ai.live.connect({
model: “models/gemini-3.5-live-translate-preview”,
config: {
responseModalities: [“AUDIO”],
translationConfig: {
targetLanguageCode: “es”,
echoTargetLanguage: false
},
inputAudioTranscription: {},
outputAudioTranscription: {}
}
})
```
-–
## Expected behavior
All spoken input (any source language) should be translated to Spanish audio output.
## Actual behavior
Inconsistent results within the same session and across sessions:
| Input (spoken English) | Expected Output | Actual Output |
|------------------------|-----------------|---------------|
| “Let’s get out of here” | “Vamos a salir de aquí” | Spanish (correct) 
|
| “I want to go to the cinema” | “Quiero ir al cine” | “I want to go to the” 
English |
| “Because” | “Porque” | Spanish (correct) |
| “Oh yes” | “Ah sí” | “Oh, yes” English |
**Reproduction rate:** ~50% of utterances output English instead of Spanish
-–
## Steps to reproduce
1. Create ephemeral token with `liveConnectConstraints` including `translationConfig.targetLanguageCode: “es”`
2. Connect via SDK with matching config
3. Speak multiple English sentences
4. Observe that some responses are in English instead of Spanish
-–
## Additional observations
- Input transcription sometimes shows wrong language (English spoken, transcribed as German)
- The `targetLanguageCode: “es”` is confirmed in logs but ignored by model
- Behavior is non-deterministic - same phrase may work in one session, fail in another