Avaya Conversational Intelligence (ACI) is a solution for transcribing, analyzing, and extracting actionable insights from millions of customer calls in real time. ACI employs multiple NLP/SLU (Spoken Language Understanding) algorithms on multilingual human-human conversations. Processing of spontaneous conversations imposes unique challenges.

Downstream NLP tasks such as intent discovery, named entity recognition, punctuation, true casing, key phrases spotting, have to take into consideration not only speech recognition errors, but also particularities of conversational speech. The mismatch between NLP models trained on written language, with correct grammar and clear sentence boundaries, and the disfluencies of spontaneous speech only exacerbates the problem.

During our presentation, we want to cover both the technical side of scaling NLP/SLU solutions to the production environment (handling multiple parallel calls from customers with low latency and high throughput), as well as scientific challenges posed by spontaneous conversations and some details of our solutions.