Azur language identification is not detecting language perfectly.

Kashish 0

I am using Azure Continuous Language Identification in Azure Speech SDK with three languages: en-IN, hi-IN, mr-IN. It behaves differently, as when I say something in English, it transcribes it into Hindi or Marathi. In short, detecting the wrong language.

this is the code i am using

var enLanguageConfig = SpeechSDK.SourceLanguageConfig.fromLanguage("en-US"); var frLanguageConfig = SpeechSDK.SourceLanguageConfig.fromLanguage("fr-FR", "The Endpoint Id for custom model of fr-FR"); var autoDetectSourceLanguageConfig = SpeechSDK.AutoDetectSourceLanguageConfig.fromSourceLanguageConfigs([enLanguageConfig, frLanguageConfig]);

santoshkc 10,140 Reputation points Microsoft Vendor

2024-11-21T14:45:39.15+00:00

Hi @Kashish,

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

It appears you're encountering issues with Azure's Continuous Language Identification in the Speech SDK, where spoken English is sometimes transcribed as Hindi or Marathi. This misidentification can occur due to similarities between languages, especially when phonetic patterns overlap or when multiple languages are mixed in the speech. The root cause might also lie in the language configurations you're using. In your current setup, you're using en-US and fr-FR, which are incorrect for your use case since you're working with en-IN, hi-IN, and mr-IN. To address this, I recommend updating the language configurations in your code to reflect the correct regional identifiers for Indian languages—specifically en-IN, hi-IN, and mr-IN. Additionally, consider reviewing your speech input to ensure it's clear and free of overlapping words across languages. If the issue persists, fine-tuning the models or adjusting for regional dialects may also help improve language detection accuracy.

To improve language detection accuracy, consider uploading an audio file instead of using real-time speech. This approach provides better control over audio quality and clarity, potentially leading to more accurate language identification and transcription.

I hope you understand! Thank you.
Kashish 0 Reputation points

2024-11-22T05:14:52.2133333+00:00

But my project is based on real-time functionality, and I was using language en-US, mr-IN, and hi-IN, which I didn't mention here. It is transcribing English into Marathi and Hindi languages. How can I improve its accuracy?
here is the example वेट विल बी 100 केजी?

Share via

Azur language identification is not detecting language perfectly.

Your answer