The v1 real-time STT endpoint returns a constant Confidence value of 0.039347406 for all NBest results globally.

Question

The v1 real-time STT endpoint returns a constant Confidence value of 0.039347406 for all NBest results globally.

OneReachAdmin-8533 0

The v1 real-time Speech-to-Text endpoint (/speech/recognition/{mode}/cognitiveservices/v1) returns a fixed Confidence value of 0.039347406 in the NBest array for every recognition result, regardless of audio quality, region, resource type, or SDK version. Text recognition is accurate — only the Confidence field is broken.

Affected endpoint

https://{region}.stt.speech.microsoft.com/speech/recognition/{mode}/cognitiveservices/v1

All modes affected: conversation, dictation, interactive

To Reproduce

curl -X POST \
  'https://eastus.stt.speech.microsoft.com/speech/recognition/conversation/cognitiveservices/v1?language=en-US&format=detailed' \
  -H 'Ocp-Apim-Subscription-Key: <ANY_VALID_KEY>' \
  -H 'Content-Type: audio/wav' \
  --data-binary @clear_speech.wav

v1 Response (broken):

{
  "RecognitionStatus": "Success",
  "DisplayText": "Hello, I'd love to help you order a pizza. What type of pizza would you like?",
  "NBest": [
    {
      "Confidence": 0.039347406,
      "Lexical": "hello i'd love to help you order a pizza what type of pizza would you like"
    },
    {
      "Confidence": 0.039347406,
      "Lexical": "hello i'd loved to help you order a pizza what type of pizza would you like"
    }
  ]
}

Note: both NBest hypotheses have identical Confidence, which should not occur for different hypotheses.

Testing performed

Variable	Values tested	v1 Confidence
Region	westus2, eastus	0.039347406
Resource type	CognitiveServices.S0 (multi-service), SpeechServices.F0 (dedicated)	0.039347406
Subscription key	key1, key2	0.039347406
SDK version	1.40.0, 1.49.0	0.039347406
No SDK (raw curl)	REST API directly	0.039347406
SpeechConfig method	fromSubscription, fromEndpoint, fromHost	0.039347406
enableDictation	on, off	0.039347406
Recognition mode	conversation, dictation, interactive	0.039347406
Output format	Simple, Detailed	0.039347406
wordLevelTimestamps	on, off	0.039347406
profanity	masked, raw	0.039347406
lidEnabled	true, false	0.039347406
initialSilenceTimeoutMs	default, 5000	0.039347406
storeAudio	true, false	0.039347406
Audio	Clean 24kHz 16-bit PCM TTS-generated WAV	0.039347406

All combinations return the same broken confidence. The Fast Transcription API returns correct confidence (0.986) for the same audio.

Expected behavior

NBest Confidence should vary based on recognition quality (typically 0.8–0.97 for clear speech). Different NBest hypotheses should have different confidence values.

Actual behavior

Confidence is pinned at exactly 0.039347406 for every NBest entry, every request, across all regions and resource types tested. The value is identical for all hypotheses and does not change with audio content — short words ("yes"), phrases ("hello how are you"), and long sentences all return the same value.

Environment

Speech SDK: 1.49.0 (also tested 1.40.0)
Also reproduced via raw REST API (no SDK)
Regions tested: eastus, westus2
Resource types tested: SpeechServices F0, CognitiveServices S0
Runtime: Node.js v20.19.6 on Ubuntu 22.04
Date first observed: April 23, 2026
Last known good: March 24, 2026 (confidence was 0.88 on same subscription)

SRILAKSHMI C 18,035 Reputation points Microsoft External Staff Moderator

2026-04-23T17:50:50.19+00:00
Hello OneReachAdmin-8533

Thank you for your detailed report and for the extensive validation.

You are absolutely correct in your observation.

We can confirm that the v1 real-time Speech-to-Text REST endpoint:

/speech/recognition/{mode}/cognitiveservices/v1?format=detailed

is currently returning a constant confidence value (0.039347406) for all entries in the NBest array.

This occurs regardless of audio quality, region, SDK version, or configuration It affects all recognition modes (conversation, dictation, interactive) It reproduces via both SDK and direct REST API calls Other APIs (e.g., batch transcription) return correct confidence values

This behavior has been identified as a service-side regression in the v1 real-time STT pipeline, specifically impacting the confidence scoring output.

The issue has been confirmed internally

It has been raised with the engineering team

The team is actively working on a fix

At this time, there is no ETA, but this is being tracked as a regression.

Recommended Workarounds

While the fix is in progress, please consider the following alternatives:

1. Use Fast Transcription API

This API returns accurate confidence scores and is recommended where real-time processing is not required.

Endpoint:

POST https://{region}.api.cognitive.microsoft.com/speechtotext/transcriptions:transcribe?api-version=2024-11-15

Sample:

curl -X POST \

Returns expected confidence values

2. Use Real-Time Streaming (v2 WebSocket Endpoint)

For real-time scenarios, we recommend switching to the v2 WebSocket-based endpoint, which is not impacted by this issue.

Endpoint:

wss://<region>.stt.speech.microsoft.com/speech/recognition/conversation/cognitiveservices/v2

Example (SDK):

const speechConfig = SpeechSDK.SpeechConfig.fromEndpoint(

Provides valid Confidence values in NBest, Suitable for real-time applications

3. Temporary Mitigation (v1 endpoint)

If you must continue using the v1 endpoint:

Treat the Confidence field as unreliable/invalid

Use alternative indicators such as:

NBest ranking order

Application-level scoring logic

Please refer this

https://learn.microsoft.com/azure/ai-services/speech-service/speech-to-text (Speech-to-Text overview)

https://learn.microsoft.com/azure/ai-services/speech-service/fast-transcription-create (Fast Transcription API)

https://learn.microsoft.com/azure/cognitive-services/speech-service/speech-sdk (Speech SDK guide)

https://learn.microsoft.com/azure/ai-services/speech-service/display-text-format#detailed-output (Detailed output/NBest format)

I hope this will help you. Please feel free to let me know if you have any other queries.

Thank you!
SRILAKSHMI C 18,035 Reputation points Microsoft External Staff Moderator

2026-04-24T17:22:57.3+00:00

Hello OneReachAdmin-8533

Did you get any chance to review the above response. Do let me know if you have any further queries.
Thank you!
Anshika Varshney 9,985 Reputation points Microsoft External Staff Moderator

2026-04-30T05:18:35.8466667+00:00

Hello OneReachAdmin-8533,

Just checking back to see if you’re still facing the same issue. If the problem persists, please share a few more details and we’ll be happy to help you further.
Thankyou!