Difference in results for Batch speech to text in speech studio and API

Question

Difference in results for Batch speech to text in speech studio and API

Sidharth Ajul 40

Hi, i was running a dual channel hindi audio file on speech studio as well as api on batch speech to text, but i am getting different results in both, and it is consistent. The speech studio version is the correct one as it recognises the phrases separately in a channel, while in the api( for which i took the parameters from speech studio as well) it is giving all the recognised phrases together instead of separate, for example, in a conversation: Channel_1: "Hello" Channel_0: "Hi" Channel_1: "How are you" Channel_0: "Im good thanks". It comes like this in speech studio but in api, it comes as: Channel_1: "Hello How are you" Channel_0: "Hi Im good thanks". I tried to combine it with the offset but still not working, the results are clearly different for speecg studio and the api. Also i tried with both, v3.2 as well as the new api-version=2024-11-15.

Speech studio output-

  "recognizedPhrases": [
    {
      "recognitionStatus": "Success",
      "channel": 0,
      "offset": "PT1.71S",
      "duration": "PT0.36S",
      "offsetInTicks": 17100000,
      "durationInTicks": 3600000,
      "durationMilliseconds": 360,
      "offsetMilliseconds": 1710,
      "nBest": [
        {
          "confidence": 0.6302351,
          "lexical": "हैलो",
          "itn": "हैलो",
          "maskedITN": "हैलो।",
          "display": "हैलो।",
          "words": [
            {
              "word": "हैलो",
              "offset": "PT1.71S",
              "duration": "PT0.36S",
              "offsetInTicks": 17100000,
              "durationInTicks": 3600000,
              "durationMilliseconds": 360,
              "offsetMilliseconds": 1710,
              "confidence": 0.6302351
            }
          ],
          "displayWords": [
            {
              "displayText": "हैलो।",
              "offset": "PT1.71S",
              "duration": "PT0.36S",
              "offsetInTicks": 17100000,
              "durationInTicks": 3600000,
              "durationMilliseconds": 360,
              "offsetMilliseconds": 1710
            }
          ]
        },
        {
          "confidence": 0.6302351,
          "lexical": "हैलो",
          "itn": "हैलो",
          "maskedITN": "हैलो",
          "display": "हैलो",
          "words": [
            {
              "word": "हैलो",
              "offset": "PT1.71S",
              "duration": "PT0.36S",
              "offsetInTicks": 17100000,
              "durationInTicks": 3600000,
              "durationMilliseconds": 360,
              "offsetMilliseconds": 1710,
              "confidence": 0.6302351
            }
          ]
        },
        {
          "confidence": 0.6302351,
          "lexical": "हैलो",
          "itn": "हैलो",
          "maskedITN": "हैलो",
          "display": "हैलो",
          "words": [
            {
              "word": "हैलो",
              "offset": "PT1.71S",
              "duration": "PT0.36S",
              "offsetInTicks": 17100000,
              "durationInTicks": 3600000,
              "durationMilliseconds": 360,
              "offsetMilliseconds": 1710,
              "confidence": 0.6302351
            }
          ]
        },
        {
          "confidence": 0.6302351,
          "lexical": "हैलो",
          "itn": "हैलो",
          "maskedITN": "हैलो",
          "display": "हैलो",
          "words": [
            {
              "word": "हैलो",
              "offset": "PT0.23S",
              "duration": "PT0.16S",
              "offsetInTicks": 2300000,
              "durationInTicks": 1600000,
              "durationMilliseconds": 160,
              "offsetMilliseconds": 230,
              "confidence": 0.6302351
            }
          ]
        },
        {
          "confidence": 0.37863052,
          "lexical": "हैलो हैलो",
          "itn": "हैलो हैलो",
          "maskedITN": "हैलो हैलो",
          "display": "हैलो हैलो",
          "words": [
            {
              "word": "हैलो",
              "offset": "PT0.23S",
              "duration": "PT0.16S",
              "offsetInTicks": 2300000,
              "durationInTicks": 1600000,
              "durationMilliseconds": 160,
              "offsetMilliseconds": 230,
              "confidence": 0.0028616383
            },
            {
              "word": "हैलो",
              "offset": "PT1.71S",
              "duration": "PT0.36S",
              "offsetInTicks": 17100000,
              "durationInTicks": 3600000,
              "durationMilliseconds": 360,
              "offsetMilliseconds": 1710,
              "confidence": 0.7543994
            }
          ]
        }
      ]
    }

Api output -

  "recognizedPhrases": [
    {
      "recognitionStatus": "Success",
      "channel": 0,
      "offset": "PT1.72S",
      "duration": "PT14.72S",
      "offsetInTicks": 17200000,
      "durationInTicks": 147200000,
      "durationMilliseconds": 14720,
      "offsetMilliseconds": 1720,
      "nBest": [
        {
          "confidence": 0.5169891,
          "lexical": "हैलो हैलो जी जी हैलो मैडम हैलो",
          "itn": "हैलो हैलो जी जी हैलो मैडम हैलो",
          "maskedITN": "हैलो हैलो जी जी हैलो मैडम हैलो",
          "display": "हैलो हैलो जी जी हैलो मैडम हैलो?",
          "displayWords": [
            {
              "displayText": "हैलो",
              "offset": "PT1.72S",
              "duration": "PT0.32S",
              "offsetInTicks": 17200000,
              "durationInTicks": 3200000,
              "durationMilliseconds": 320,
              "offsetMilliseconds": 1720
            },
            {
              "displayText": "हैलो",
              "offset": "PT4.52S",
              "duration": "PT0.52S",
              "offsetInTicks": 45200000,
              "durationInTicks": 5200000,
              "durationMilliseconds": 520,
              "offsetMilliseconds": 4520
            },
            {
              "displayText": "जी",
              "offset": "PT8.12S",
              "duration": "PT0.48S",
              "offsetInTicks": 81200000,
              "durationInTicks": 4800000,
              "durationMilliseconds": 480,
              "offsetMilliseconds": 8120
            },
            {
              "displayText": "जी",
              "offset": "PT11.28S",
              "duration": "PT0.44S",
              "offsetInTicks": 112800000,
              "durationInTicks": 4400000,
              "durationMilliseconds": 440,
              "offsetMilliseconds": 11280
            },
            {
              "displayText": "हैलो",
              "offset": "PT14.32S",
              "duration": "PT0.52S",
              "offsetInTicks": 143200000,
              "durationInTicks": 5200000,
              "durationMilliseconds": 520,
              "offsetMilliseconds": 14320
            },
            {
              "displayText": "मैडम",
              "offset": "PT14.84S",
              "duration": "PT0.24S",
              "offsetInTicks": 148400000,
              "durationInTicks": 2400000,
              "durationMilliseconds": 240,
              "offsetMilliseconds": 14840
            },
            {
              "displayText": "हैलो?",
              "offset": "PT15.92S",
              "duration": "PT0.52S",
              "offsetInTicks": 159200000,
              "durationInTicks": 5200000,
              "durationMilliseconds": 520,
              "offsetMilliseconds": 15920
            }
          ]
        }

Sidharth Ajul 40 Reputation points

2025-02-13T12:39:48.4366667+00:00
Update: Fixed it found out it was using different models for transcription, the speech studio was using the model 20241211 -

dc55170a-871c-4747-886c-39c385f30e38

and api was using the moel 20240926 Batch Transcription

22b2ceb8-baea-4e34-872c-db1dedbf0eef

But i still dont know why my api didnt select the latest model id when i didnt specifically mention any model id. Can anyone explain this?

Accepted answer

0 additional answers

Your answer

Sidharth Ajul 40 Reputation points

2025-02-13T12:39:48.4366667+00:00

Update: Fixed it found out it was using different models for transcription, the speech studio was using the model 20241211 -

dc55170a-871c-4747-886c-39c385f30e38

and api was using the moel 20240926 Batch Transcription

22b2ceb8-baea-4e34-872c-db1dedbf0eef

But i still dont know why my api didnt select the latest model id when i didnt specifically mention any model id. Can anyone explain this?

Answer 1

Prashanth Veeragoni 1,755 Microsoft External Staff

Hi Sidharth Ajul,

Thank you for reaching out to Microsoft Q&A forum!

It looks like the root cause of the issue was that different models were used for transcription in Speech Studio and the API.

Speech Studio used model dc55170a-871c-4747-886c-39c385f30e38 (2024-12-11) and API used model 22b2ceb8-baea-4e34-872c-db1dedbf0eef (2024-09-26 Batch Transcription).

The issue was caused by the API using an older model (20240926 Batch Transcription) while Speech Studio automatically selected the latest model (20241211). Since Speech Studio dynamically assigns the best available model, but the API defaults to an older model unless specified, you experienced different transcription behaviours.

Azure Speech API does not always default to the latest model unless explicitly specified. It typically selects the most stable version unless overridden by the user.

To Fix this issue Specify the Latest Model in API Calls:

To ensure the API uses the same model as Speech Studio, explicitly specify the latest model (20241211) in your API request. You can do this by adding the model parameter in your API request.

Modify your batch transcription request JSON by including the latest model ID (dc55170a-871c-4747-886c-39c385f30e38).

Re-run your API request and compare results with Speech Studio.

Please refer below document - About Batch Transcription
https://learn.microsoft.com/en-us/azure/ai-services/speech-service/batch-transcription

Hope this helps. Do let us know if you any further queries.

-------------

If this answers your query, do click Accept Answer and Yes for was this answer helpful.

Thank you.

Sidharth Ajul 40 Reputation points

2025-02-14T07:18:39.6766667+00:00

Thanks for the answer, also is there a page where i can get all the latest model id's for each language, instead of checking for each language individually.
Deleted

This comment has been deleted due to a violation of our Code of Conduct. The comment was manually reported or identified through automated detection before action was taken. Please refer to our Code of Conduct for more information.
Prashanth Veeragoni 1,755 Reputation points Microsoft External Staff

2025-02-14T13:51:58.97+00:00
Hi Sidharth Ajul,

While Microsoft does not maintain a dedicated public page for model IDs

But you can get Latest Model IDs for Each Language by following below steps:

Azure provides an API endpoint to list the available models for different languages. You can retrieve this information dynamically using the following API call:

Call the Get Models API: Use the following REST API request to fetch all available models:

GET https://<region>.api.cognitive.microsoft.com/speechtotext/v3.1/models Ocp-Apim-Subscription-Key: YOUR_AZURE_SPEECH_KEY

The response will include a list of available models, their IDs, and supported languages:

{ "values": [ { "modelId": "dc55170a-871c-4747-886c-39c385f30e38", "kind": "Transcription", "locale": "en-US", "createdDateTime": "2024-12-11T00:00:00Z" }, { "modelId": "22b2ceb8-baea-4e34-872c-db1dedbf0eef", "kind": "Transcription", "locale": "fr-FR", "createdDateTime": "2024-09-26T00:00:00Z" } ] }

The modelId field represents the ID of each available model.

The locale field indicates the language.

The createdDateTime helps identify the latest model.

Hope these answers your query.

Thank You.

Share via

Difference in results for Batch speech to text in speech studio and API

0 additional answers

Your answer