Bug: Intermittent ASR Failures and Incorrect Language Transcription for Vietnamese Input

**Title:** Bug: Intermittent ASR Error and Incorrect Language Transcription for Vietnamese Input

**Message:**

Hello Gemini Team,

We are experiencing serious issues with the Automatic Speech Recognition (ASR) component of Live API when using Gemini API for Vietnamese input. These issues have started occurring over the past week; the system was working fine before this point.

### Problem Description

There are two main issues:

1. **Intermittent empty ASR response:** The ASR service occasionally fails to return any transcribed text from Vietnamese audio input. Sometimes it provides an incorrect transcription, but other times it hangs and does not return. **This actually affects the generative model (e.g. `gemini-live-2.5-flash-preview`), which cannot run without transcribed text. This causes the interaction to stall, making the app unusable.**

2. **Incorrect language transcription:** Even when ASR returns text, it sometimes mistranslates Vietnamese speech into completely different languages, such as Thai or Chinese. Our ASR is explicitly configured for Vietnamese (`vi-VN`).

This appears to be a regression, as we did not encounter these issues about two weeks ago with a similar setup.

### Steps to reproduce

1. Set up an environment with Python 3.12 and `google-genai==1.45.0`.
2. Initialize a generative model session using `gemini-live-2.5-flash-preview` or `gemini-2.0-flash-live-001`.

3. Continuously stream Vietnamese audio input to the API's ASR service.

4. Observe the transcribed text returned by ASR.

5. This issue occurs intermittently, so it may take multiple attempts to observe both the blank response and the incorrect language transcript.

### Expected Behavior

* ASR must consistently and reliably convert Vietnamese audio data to Vietnamese text.

* ASR must never convert Vietnamese audio to Thai, Chinese, or any other language when configured for Vietnamese.

* In cases where the audio is silent or unclear, ASR must return a predictable empty string without causing the main model call to hang.

### Actual Behavior

* The ASR service is unreliable, sometimes returning text and sometimes nothing for similar audio inputs.
* Downstream Gemini model calls stop responding when ASR does not provide text.

* ASR often returns incorrect language records (Thai, Chinese).

### Environment Details

* **Python Version:** 3.12
* **Library Version:** `google-genai==1.45.0`
* **Models Tested:** `gemini-live-2.5-flash-preview`, `gemini-2.0-flash-live-001`
* **Input Language:** Vietnamese
* **ASR Configuration:** Set to Vietnamese (`vi-VN`)

The core issue seems to be in the ASR service layer rather than the generative models themselves. Since this phenomenon has only recently appeared, we suspect a recent service update may be the cause.

Please let us know if you need any further information or logs from us. Thank you for bringing this important matter to our attention.