Bug: Flash 2.5 transcription exhausts output tokens due to repeated `[unclear]`, resulting in truncated JSON (recent regression)

[ish-codes-magic]

Describe the bug
When using Flash 2.5 on Vertex AI for audio transcription with the google-genai package with batching enabled, the model repeatedly outputs the literal token [unclear]. This repetition consumes the entire max_output_tokens budget before transcription completes, causing the response to be truncated and resulting in invalid or incomplete JSON.

This behavior appears to be a recent regression. The same transcription pipeline was significantly more reliable approximately 1–1.5 months ago, with far fewer [unclear] repetitions and successful completion of JSON responses.

---

Environment

• Platform: Vertex AI
• Model: Flash 2.5
• Library: google-genai
• Task: Audio transcription with batching
• Response MIME type: application/json
• Response schema: Enabled
• Thinking mode: Diabled

---

Steps to reproduce

1. Send an audio file via file_uri with a transcription prompt
2. Enable structured JSON output using response_schema
3. Set max_output_tokens appropriate for the expected transcription length
4. Invoke Flash 2.5 on Vertex AI with batching

---

Expected behavior

• The model should avoid excessive repetition of [unclear]
• The model should complete transcription within the token budget
• The model should consistently return a valid JSON response conforming to the schema

---

Actual behavior

• The model repeatedly emits [unclear] segments
• Output tokens are exhausted before transcription completes
• JSON output is truncated or malformed

---

Code snippet

parts = [
    {
        "file_data": {
            "file_uri": uri,
            "mime_type": self._get_mime_type(file_path)
        }
    },
    {
        "text": final_prompt
    }
]

generation_config = {
    "response_mime_type": "application/json",
    "temperature": transcription_config.TEMPERATURE,
    "max_output_tokens": transcription_config.get_max_output_tokens(model),
}

schema_class = get_transcription_result_class(model, phase)
if schema_class:
    if isinstance(schema_class, dict):
        generation_config["response_schema"] = schema_class
    else:
        schema_dict = schema_class.model_json_schema()
        schema_dict = self._resolve_json_schema_refs(schema_dict)
        schema_dict.pop("$defs", None)
        generation_config["response_schema"] = schema_dict

generation_config["thinking_config"] = {
    "thinking_budget": transcription_config.THINKING_BUDGET
}

instance = {
    "id": str(i - 1),
    "request": {
        "contents": [
            {
                "role": "user",
                "parts": parts
            }
        ],
        "generation_config": generation_config
    }
}

---

Additional context

• Increasing max_output_tokens doesn't reduce the issue.
• The regression has been observed consistently over the last 1–1.5 months

---

Questions

• Is this a known regression in Flash 2.5 transcription behavior?
• Are there recommended mitigations to prevent token exhaustion due to repeated [unclear] output?
• Is Flash 2.5 currently recommended for transcription workloads on Vertex AI?

[Venkaiahbabuneelam]

Hi @ish-codes-magic , Thanks for reaching out us!
We tried to reproduce the issue using same configs you mentioned and provided audio file for transcription but we were ble to get the response as expected.

Thanks

[github-actions]

This issue has been marked as stale because it has been open for 7 days with no activity. It will be closed in 2 days if no further activity occurs.

[github-actions]

This issue was closed because it has been inactive for 9 days. Please post a new issue if you need further assistance. Thanks!