Nova, Explained Simply

Deepgram's Nova is a speech-to-text model, which means it listens to people talking and turns that audio into written text. Anyone who builds apps that rely on voice—like video captioning tools, meeting transcribers, or language-learning apps—would want to use this to get highly accurate written words out of spoken audio.

What it does well

• Highly versatile: Instead of only being good at understanding one specific type of audio (like just phone calls or just news broadcasts), it performs well across a wide variety of contexts.
• Deeply trained: It was trained on 47 billion "tokens" (which are basically small chunks of words and sounds) across more than 100 different topics. This massive training helps it understand diverse voices and speaking styles accurately.
• Multilingual: It does a great job with English and now supports Spanish, including both the Spanish spoken in Spain (es) and Latin America (es-419).

How it compares

• Versus older Deepgram models: Nova is Deepgram's most powerful and affordable speech-to-text model released to date. Unlike previous models that might have been built for specific tasks, Nova is designed to be a highly accurate all-rounder across many different topics.
• Versus rivals: According to Deepgram, Nova is the "deepest-trained" automatic speech recognition (ASR) model available right now, meaning it has been fed a wider variety of training data than competing models.

What it costs and what it can handle

• Cost: It is priced to be Deepgram's most affordable model.
• Language support: It can handle English and Spanish (specifically the es and es-419 dialects).

Worth knowing

• You can try it out right now without writing any code by using Deepgram's "API Playground" (a testing sandbox on their website).
• While it supports a massive variety of topics, it currently only handles English and Spanish, so it won't work for other languages just yet.