Version 11.3

v11.3.1-7

ENHANCEMENTS

Common Changes

• None.

REST Changes

• None.

MRCP Changes

• None

IMPROVEMENTS

• Support of quantized language models. A little faster and less ram usage compared with the previous E2E models.
• Updated onnxruntime to v1.14.1.

KNOWN ISSUES

• None.

v11.3.1-1

ENHANCEMENTS

Common Changes

• None.

REST Changes

• None.

MRCP Changes

• Added MRCP enable/disable configuration.

IMPROVEMENTS

• Fixed confidence computation for http recognitions.

KNOWN ISSUES

• When Offline Audio Splitter is set to "realtime-var", the utterance in confidence value in Nbestlist may not be calculated correctly.

v11.3.0-5

ENHANCEMENTS

Common Changes

• Support Dolphin models: Our platform now supports Dolphin models, which we call E2E (next-gen language models) allowing to make recognition with higher recognition rates.
• Added language information to models: We've added language information to our models, making it easier to select the right model for language needs.
• Added “LowercaseResultsInTurkish” config.
• Support multiple allowed origins: Specify multiple allowed origins for API calls, providing more flexibility and control over how the system is accessed.
• Added log count statistics: Allowing to monitor and analyze system's performance and usage.
• Updated ONNX Runtime to v1.14.1.

REST Changes

• Added REST endpoint to get model default versions: Retrieving the default version of a model using a new REST endpoint is possible.
• Separated SR and management REST ports: We've separated the speech recognition and management REST ports, providing better security and control over access to the system.
• Removed HTTP tests from smoke tests.

WebSocket Changes

• Added WebSocket test to /web/doc/ws folder: We've added a new WebSocket test to our /web/doc/ws folder, allowing to test WebSocket connections and troubleshoot any issues.

MRCP Changes

• Added MRCP enable/disable configuration.

IMPROVEMENTS

• The problem of not being able to recognize Persian and Pashto models via LMS has been solved.

KNOWN ISSUES

• If recognition is made with a very long audio file that exceeds the maximum character limit of 7000 for the SQL server, the recognition result is not written to the "Recognitions" page.
• Masking feature does not work as a result of grammar recognition with REST.
• The "Mrcp Bulk Dictate" scenario fails in the smoke test. This is normal as the models required for testing are not available in the customer environment.
• SRMangal also fails the "Http Dictate Bulk" scenario. This is normal as the models required for testing are not available in the customer environment.
• Dolphin models (E2E) do not support nbest list. There may be errors related to this in the logs.

LMS - Language Model Service

v11.3.0-5

ENHANCEMENTS

• None.

IMPROVEMENTS

• Updated onnxruntime to v1.14.1 to support quantized E2E language models.

KNOWN ISSUES

• None.

v11.3.0-2

ENHANCEMENTS

• Added Farsi and Pashto languages.
• Added language information to HTTP models response.
• Support Dolphin models (only hosting is supported, can't train Dolphin models yet).

IMPROVEMENTS

• None.

KNOWN ISSUES

• If a header containing Turkish characters is added in the header "CustomResponseHeaders" settings in the LMS AppSettings.json file, "Invalid non-ASCII or control character in header" is written in the LMS logs. In this case, the requested headers will not be added.
• Pronunciation of the word cannot be added to the Arabic language model via LMS Explorer. (It is certain that it cannot be added with copy paste, but it is unknown whether it can be added by hand.)
• There is no health check method in this version.
• An “Unauthorized" error is received when trying to train the model via LMS Explorer while there is no LMS license.
• When the "LANGUAGE MODELS" and "LISTENERS" tabs are pressed on the -localhost:50001 page, it is not possible to switch between pages. The information is displayed on the same page.