A recent study by AI4Bharat highlights significant challenges faced by global AI models like OpenAI, Meta, and Microsoft in accurately transcribing Indian languages. OpenAI's models report a Word Error Rate (WER) exceeding 55% in Indian speech, with some dialects, such as Bhojpuri and Maithili, falling to 66% errors. In stark contrast, Indian startup Sarvam achieves over 93% accuracy across multiple languages, demonstrating the effectiveness of localized solutions in speech recognition.
A recent study by AI4Bharat highlights significant challenges faced by global AI models like OpenAI, Meta, and Microsoft in accurately transcribing Indian languages. OpenAI's models report a Word Error Rate (WER) exceeding 55% in Indian speech, with some dialects, such as Bhojpuri and Maithili, falling to 66% errors. In stark contrast, Indian startup Sarvam achieves over 93% accuracy across multiple languages, demonstrating the effectiveness of localized solutions in speech recognition.