Share via

Azure Speech Recognition: noise or chatter is recognized as insurance related texts, rather that just not recognized.

Gijsbert Huijsen 21 Reputation points
2025-07-24T10:16:59.59+00:00

Hello,

When noise or background chatter is sent to Azure Speech Recognition it sends back texts related to insurance policies. Texts which where not spoken at all.

For example repsonse like (for language nl-NL):

ik wil graag een verzekering afsluiten voor mijn hypotheek
ik wil graag weten of mijn verzekering
ik wil graag weten of mijn verzekering het risico op mijn verzekeringspolis

English translation for the examples:
I would like to take out insurance for my mortgage.
I would like to know if my insurance
I would like to know if my insurance covers the risk on my insurance policy.

These sentences are confusing our users if they occur in the middle of a translation about a completely different subject.

I had audio logging enabled, so I have the logs (json + wav audio files) where this occurs. I don't want to publish those audio files publicly, but I can supply them if needed.

Can you adjust the speech recognition to just return an empty recognition result and not default to texts about insurances if Azure cannot recognize the sound?

Thanks,

Gijsbert Huijsen

Azure AI Speech
Azure AI Speech
An Azure service that integrates speech processing into apps and services.
Accepted answer
  1. Pavankumar Purilla 10,350 Reputation points Microsoft External Staff Moderator
    2025-07-24T13:10:32.97+00:00

    Hi Gijsbert Huijsen,

    The behavior you’re observing where Azure Speech Recognition returns unrelated insurance-related text in response to background noise or chatter—is a result of the service attempting to interpret unclear or unintelligible audio using learned language patterns. This can lead to inaccurate transcriptions that were never actually spoken, which we understand can be confusing for users. While the service does not currently support a built-in option to return an empty result for unrecognized input, there are several ways to improve accuracy. First, we recommend checking the acoustic environment to ensure it's as quiet as possible, using high-quality microphones, and optimizing their placement to reduce external noise. Additionally, you can fine-tune recognition behavior by adjusting SpeechConfig parameters such as increasing the Initial Silence Timeout to allow more time before speech detection begins, and modifying the Segmentation Silence Timeout to better handle pauses between phrases, reducing the risk of misinterpreting background sounds. We also suggest enabling word-level confidence scores or retrieving nBest alternatives, which allows your application to detect and filter out low-confidence transcriptions.

    1 person found this answer helpful.

1 additional answer

Sort by: Most helpful
Most helpful Newest Oldest
  1. Gijsbert Huijsen 21 Reputation points
    2025-07-25T08:14:17.48+00:00

    Hi @Pavankumar Purilla ,

    Thanks for your answer. We will try to use the confidence scores that you mentioned. Thanks for the tip!

    I don’t understand why noise always results in texts about insurance policies, I would expect random words or different random sentences, not always the same sentences.

    But we will try the confidence scores to filter out low confidence results.

    Thank you for the information.

    Kind regards,

    Gijsbert

    0 comments

Your answer

Answers can be marked as Accepted Answers by the question author, which helps users to know the answer solved the author's problem.