Elevated error rate (>95%) with `gemini-1.5-flash` when processing audo

I noticed that Gemini API returns errors consistently when the prompt contains audio data. This started happening on 2025-04-02 according to my API usage stats.

Error Messages

The error is one of these two:

  1. [GoogleGenerativeAI Error]: Error fetching from https://generativelanguage.googleapis.com/v1beta/models/gemini-1.5-flash-002:streamGenerateContent?alt=sse: [503 Service Unavailable] The model is overloaded. Please try again later.
  2. [GoogleGenerativeAI Error]: Error fetching from https://generativelanguage.googleapis.com/v1beta/models/gemini-1.5-flash-002:streamGenerateContent?alt=sse: [500 Internal Server Error] Unable to submit request because the service is temporarily unavailable.

Google Cloud API Metrics

In Google Cloud Console → APIs & Services → Generative Language API, as of writing this post the error rate is over 95% making the service almost unusable. Billing is turned on and I am using the paid API, not the free tier.

Gemini API Status

Despite the very high error rate, Gemini API’s status page says that “All Systems Operational.” :face_with_diagonal_mouth:

I suggest that Gemini team implement a separate status check for audio processing.

1 Like

Update: I noticed that Gemini’s error rate when processing audio seems to be very high in just some regions (such as US) but not the others. So I set up a VPS in a different region to call the Gemini API from there, and the error rates improved.

Hi @dtinth,

Thanks for reporting the issue! I’ll share it with the team. Also, if you consistently use a single audio format, could you please tell me what it is

Thank you

1 Like

Hi @dtinth , had the same issue, described here: Gemini-1.5-flash is no longer processing audio files (500 Exception) - retry does not help - #5 by Siva_Sravana_Kumar_N

I saw exactly the same behaviour as you, started about a week go, with increasing error rates. Clearly an issue in the backend.

As per my message in the other thread, my workaround is switching to the flash-2.0-lite.

1 Like

All of the requests sent from our service is audio/x-m4a. Thank you.