Case Study: Dialpad achieves 88%+ transcription accuracy and scalable conversational ML training data with Appen

Dialpad Creates Data That Powers ML Models for Human Conversation at Scale

Dialpad collects telephonic audio and uses in-house speech recognition and NLP to analyze every customer conversation, but needed robust, company-specific training data to make those models accurate. After topping out at about 70% label accuracy with a previous provider, Dialpad turned to Appen for human-in-the-loop data services, using the Appen platform for audio transcription, categorization, and verification.

Appen ran jobs to transcribe audio, categorize key moments, verify model outputs, and even used geolocation tools to ensure correct labeling of U.K. idioms. Within a couple weeks Appen’s work raised labeler accuracy to about 88% (and kept it in the high 80s to 90s), enabling Dialpad to scale custom training datasets across many customers while maintaining that higher accuracy.

Dialpad

Etienne Manderscheid

Head of Data Science, Co-Founder TalkIQ