Deepgram Nova-3

Nova-3 arrived in early 2025 as Deepgram's flagship transcription engine and, unlike the voice models that get the headlines, it is the kind of software that quietly runs call centres, medical dictation and meeting notes. Deepgram's own published benchmarks put its median word error rate at 6.84 per cent on streaming audio against 14.92 per cent for the next-best competitor, and 5.26 against 10 per cent in batch mode; independent testing lands more conservatively, in the region of 88 to 92 per cent accuracy on clean English, which is roughly level with Whisper and Google's Chirp. Two features explain its position in industry. It transcribes ten languages in real time and can follow a speaker who switches between them mid-sentence — code-switching, the everyday reality of bilingual households and international call queues. And it accepts up to a hundred key terms supplied at request time, so drug names, part numbers or a company's product vocabulary can be taught to the model without retraining it. Pricing is metered by the minute, and the model can be self-hosted where audio must not leave the building.

▸Videos

No videos yet.

▸Reviews