Humyn Labs study finds AI transcription tools fail Indian languages
A new Humyn Labs study found that most AI transcription tools just don't get Indian languages right.
They often mess up audio, especially when people mix English and Indian languages (think: Hinglish).
Manish Agarwal from Humyn Labs explained it's mainly because the issue is tied to English-first, internet-trained benchmarks and limited independent validation.
Study urges testing on code-mixed speech
The research showed a major difference between what these tools claim and how they actually perform on Indian speech.
Deepgram Nova-3 topped the list for understanding meaning, while Amazon Transcribe and OpenAI's models lagged behind.
Even Sarvam AI's tool, built for Indian languages, struggled with code-mixed speech.
The takeaway? For AI to work well here, it needs to be tested on the way we actually talk, mixing languages and all.