Nova-3 Expands Speech-to-Text Support for Thai, Cantonese, Mandarin, and Indic Languages
Blog post from Deepgram
Deepgram's Nova-3 has expanded its speech-to-text transcription capabilities across the Asia-Pacific region, now supporting Thai, Cantonese Traditional, Mandarin Simplified, Mandarin Traditional, and Gujarati, while improving accuracy for Bengali, Marathi, Tamil, and Telugu. This expansion comes with substantial enhancements in transcription quality, notably reducing Word Error Rate (WER) compared to the previous Nova-2 model, with Thai achieving a 69.43% reduction and Mandarin Simplified a 65.21% reduction. The advancements address the challenges posed by tonal languages, multiple writing systems, and regional speech variations, enhancing both batch and streaming use cases essential for enterprise-grade voice AI applications. These updates are seamlessly integrated into the existing API, allowing developers to leverage the improved capabilities without additional training or configuration, and emphasize Deepgram's commitment to supporting diverse linguistic environments and regional speech patterns in customer support, conversational AI, transcription, and analytics workflows.