In February 2025, Deepgram unveiled Nova-3, the latest iteration of its popular speech-to-text (STT) model.
As customer service teams increasingly rely on AI-driven solutions, such models are becoming an increasingly crucial factor in ensuring high-quality experiences.
However, traditional STT solutions have struggled to maintain accuracy amid background noise, overlapping speech, and changing language.
To overcome these challenges, Nova-3 introduces several new capabilities, alongside increasing accuracy rates.
Indeed, Nova-3 demonstrates more accurate transcriptions in environments with loud background noise and overlapping speakers.
As such, it promises to advance the performance of AI use cases in call centers, retail, and drive-thrus where clarity and accuracy are paramount.
Yet, transcription accuracy in bustling places is just one notable feature that builds on Deepgram’s previous model: Nova-2. Others include:
- Real-Time Multilingual Transcription: Nova-3 processes conversations in multiple languages, which could enable voice translation use cases where customers and service reps to interact in near real time.
- Industry-Specific Vocabulary: Businesses can fine-tune Nova-3 to recognize and understand specialized industry terms, enhancing accuracy across sectors
- Redaction of Sensitive Information: Nova-3 removes sensitive personal information from live conversation transcriptions, which is a crucial feature for industries such as finance and healthcare.
“Nova-3 represents a significant leap forward,” claimed Scott Stephensen, CEO of Deepgram.
By integrating advanced architectural enhancements and extensive training across diverse datasets, we’ve developed a model that not only meets but exceeds the evolving needs of our clients across various industries.
These clients cross many sectors. For instance, Deepgram has found success in working with emergency services teams.
Nova-3’s ability to handle language-switching will support these teams further, ensuring vulnerable people can get the support they need, no matter the languages they speak.
Meanwhile, consider that industry-specific vocabulary feature. That’s critical in industries like healthcare, finance, and legal to ensure confidentiality and customer security.
Moreover, contact centers may enhance quality assurance and AI-driven customer insights, ensuring they have a clear, objective view of trends, performance, and customer sentiment.
Elsewhere in the contact center, Nova-3 promises to improve productivity and compliance by creating accurate and reliable call logs.
After all, real-time transcripts still prove a helpful tool in scenarios where agents are struggling with connection issues or if the call center is particularly noisy.
Yet, the applications of Nova-3 extend far beyond the contact center. Indeed, in automated ordering systems, retail stores, and beyond, this technology can positively enhance customer experiences.
Wait… Just How Accurate Is Nova-3
According to Deepgram’s assessments of Nova-3’s performance, the model has a median Word Error Rate (WER) of 6.84 percent in real-time audio situations.
That’s a significant improvement on Nova-2, which held a median WER of 8.4 percent.
Yet, it also surpasses competitive offerings like Amazon Transcribe, Microsoft Azure Batch v.31, and OpenAI.
These score a WER rate of 8.98 percent, 7.81 percent, and 8.83 percent, respectively.
Furthermore, for pre-recorded audio data, Nova-3 achieved a median WER of 5.26 percent, which it claims is a 47.4 percent improvement over a leading competitor.
“We saw a massive jump in accuracy with Nova-3,” said Brendan Chan, CTO at Talkatoo, after his company deployed the model.
Previous models recognized only 10 percent of critical veterinary terms, but with Nova-3 and keyterm prompting, we’re seeing 625 percent improvement in keyterm recognition.”
With this in mind, Nova-3 could help optimize the efficiency and output of AI models across the CX sphere.
Join the CX Community That Values Your Voice
This is your space to speak up, connect, and grow with thousands of CX leaders. Share your voice, influence what’s next, and learn from the best in customer experience. Join the conversation today.
