Want to Confirm about the Confidence score generated by Azure AI Speech

Hello @Divyansh Modi （モディ ディビヤンシュ）

The confidence score returned by Azure AI Speech is generally a recognition confidence score, indicating how confident the speech-to-text engine is in the recognized text, rather than a direct measure of the overall accuracy of the transcription.

To verify the confidence values correctly, please ensure that:

• You are using the Detailed output format, as confidence scores are available in the NBest section of the response.
• You are reading the NBest[].Confidence field and not relying solely on the displayed transcription text.
• The audio degradation is significant enough to impact recognition. Minor or artificial noise may not always cause confidence scores to drop if the speech remains understandable.

It's also important to note that confidence scores are not calibrated as "High", "Medium", or "Low" accuracy guarantees. A transcript can have a high confidence score and still contain errors, and vice versa. Therefore, confidence values should be used as an indicator of recognition certainty rather than a definitive measure of transcription quality.

If you continue to observe consistently high confidence scores across clearly corrupted or unintelligible audio samples, please share:

• A sample confidence score output
• The Speech SDK or API version being used
• Whether the transcription is generated via Speech Studio, Speech SDK, REST API, or Azure AI Foundry

This will help determine whether the behavior is expected or requires further investigation.

Thankyou!

Hi Divyansh Modi （モディ ディビヤンシュ）,

**
**Welcome to Microsoft Q&A,

Confidence scores in Azure AI Speech should range from 0.0 (no confidence) to 1.0 (full confidence), and can absolutely return low scores for poor audio quality or noise. The issue is likely in how you're retrieving or testing them.

Here's what you need to verify:

1. Use Detailed Format in REST API

Make sure you're requesting the detailed format to access confidence scores:

https://<region>.stt.speech.microsoft.com/speech/recognition/conversation/cognitiveservices/v1?language=en-US&format=detailed

2. Check NBest Array in Response

The confidence score is only in the NBest array, not the top-level response:

{
  "NBest": [
    {
      "Confidence": 0.905,  // This is the confidence score
      "Display": "What's the weather like?"
    }
  ]
}

3. Test with Known Low-Confidence Scenarios

If you still see high scores with intentional noise, check:

• Are you parsing NBest[0].Confidence correctly?
• Is your audio actually reaching the API (use cloud debugging tools)?
• Try the REST Speech to Text quick test in Speech Studio to isolate the issue

4. Known Behavior with Poor Audio

Azure Speech does return low confidence (0.11-0.15 range) for misrecognized noise. Example from real implementation: noise recognized as "S" returned Confidence: 0.11938265.

Next Steps:

• Enable detailed logging and share the actual NBest response (not just DisplayText)
• If scores remain biased high across multiple noise tests, escalate via Azure Portal support since this would indicate a model or service behavior anomaly

Reference: Speech to text REST API confidence scores

Please click Accept Answer and upvote if this helped.