Real-time conversational insights from phone call data

Transcription

Transcribing phone calls accurately presents unique challenges due to the variety of dialects and accents involved. Mike McCourt explains that while the phone calls are transcribed into text before reaching the data science team, significant differences in dialect and accent can affect the quality of these transcripts 1. The transcription process can range from nearly perfect to almost entirely noise, with words being mistranscribed due to audio quality issues, creating artificial synonyms 2. McCourt notes, "We treat it almost as like a quasilinguistic phenomenon," highlighting the complexity of dealing with such variability.

   

Call Variability

Handling the variability in phone call data requires sophisticated classification models. McCourt points out that phone calls can vary widely in length, from under a minute to over two hours, complicating the classification process 3. The freedom of expression in English adds another layer of complexity, as no two people will express themselves in exactly the same way, even when performing the same task 3. To address these challenges, Invoca's models are trained automatically on customer-uploaded data, selecting the best-performing model to ensure reliable results 4. McCourt explains, "The customer uploads the data, and the models train automatically."