Deepgram's Voice AI Now Cheaper Thanks to AWS Chips
Your voice assistant could get cheaper and faster with this tech.
Deepgram is a company that helps computers understand and generate human speech. They process a massive amount of audio—quadrillions of tokens each year—for over 200,000 developers and 1,400 organizations. To keep up with demand and reduce costs, Deepgram worked with Amazon Web Services (AWS) to optimize their speech-to-text model, Nova-3, for AWS's custom Trainium2 chips. These chips are designed specifically for AI tasks, and by tuning their model to run efficiently on them, Deepgram can now process more audio for the same amount of money.
The optimization process involved rewriting parts of the model to take full advantage of the Trainium2 hardware. Deepgram and AWS engineers used special tools to profile the model's performance and replace inefficient parts with faster, custom-built components. As a result, they achieved an 11–70% increase in audio processed per dollar compared to their previous setup. This means that for every dollar spent on computing, Deepgram can now transcribe more audio, which is especially important for high-volume tasks like analyzing customer service calls or transcribing medical records.
For everyday users, this could mean that voice AI services become cheaper and more reliable. Businesses that rely on transcription or voice assistants might see lower costs, and developers could build more advanced voice applications without breaking the bank. Additionally, customers who run Deepgram's models on their own AWS infrastructure can also benefit from these efficiency gains, making it easier and more cost-effective to deploy voice AI in their own products.
While the technical details are complex, the bottom line is simple: by making AI models run more efficiently on specialized hardware, companies like Deepgram can offer better services at lower prices. This is part of a broader trend where AI companies are increasingly turning to custom chips to handle the heavy lifting, which could lead to more powerful and affordable AI tools for everyone.
- Deepgram's speech-to-text model now runs more efficiently on AWS's Trainium2 chips, processing up to 70% more audio per dollar.
- This optimization could lead to lower prices for voice AI services used by businesses and developers.
- The improvements also benefit customers who run Deepgram's models on their own AWS infrastructure.
Why It Matters
Cheaper voice AI means more affordable transcription services and voice assistants for businesses and consumers.