June 2021 Summaries

Deepgram, a leading automatic speech recognition vendor, has experienced significant growth in Q2 2021 with 3.2x YoY revenue growth and 2.7x YoY employee growth. The company's success is driven by increasing ASR adoption across industries, expanding software customer base, and key executive hires. Deepgram recently appointed four new executives and expanded its base of software customers in the CCaaS and Call Analytics verticals. The global market for speech and voice recognition technology is projected to reach $31.82 billion by 2025.

The article discusses the challenges of achieving a human-like conversational AI experience, focusing on technical and data obstacles. It highlights the need for better transcription speed and accuracy optimization, improved NLP and NLU, enhanced text-to-speech engines, and tighter integrations between various components of the workflow. The article also emphasizes the lack of training data as a significant challenge in creating voicebots for all languages, accents, dialects, and use cases. It concludes by stating that although there is no easy button to create a perfect overall Conversational AI voicebot, single-use case voicebots are currently available, with experts predicting a big evolution in this space within the next two to three years.

Deepgram, an automatic speech recognition company, has introduced three new product features - conversational AI, sales and support enablement, and real-time streaming. These features aim to enhance human-like virtual agents and improve voice tasks like billing, support, add-on sales, compliance or ID verifications. Conversational AI reduces response lag time to less than 300 milliseconds, while the sales and support enablement feature helps reduce customer churn with faster ASR for real-time offers and alerts, and improves employee success with accurate call analytics and transcripts for coaching.

The article discusses the use of unsupervised learning for speech recognition and whether it yields more accurate results than other methods. It explains that AI training methods for speech recognition include supervised, self-supervised (unsupervised), and semi-supervised learning. Supervised learning involves labeled training data sets, while unsupervised learning does not have paired audio and text files. Semi-supervised is a combination of both. The article argues that to create speech models that can survive in the wild, a combination of supervised and unsupervised learning techniques is necessary. Deepgram has invested heavily in labeled training data and continues to improve its speech training methodologies for better model performance.