We are trying to leverage the Gong integration to identify the longest monologue by a sales rep. The way Gong tracks longest monologue is inaccurate because even if a prospect coughs, it ends the monologue. We have tried for hours using a prompt to analyze the transcript to better identify longest monologue and it's still inaccurate (but much closer to gong). Does anyone have any suggestions/ideas on how we can achieve this? (prompt in thread)
#CONTEXT# You’re finding the rep’s longest monologue in a transcript. A monologue only ends if the prospect speaks for ≥10 seconds. #OBJECTIVE# Analyze the call transcript and return the rep's longest uninterrupted monologue, following the specified rules. #INSTRUCTIONS# 1. Analyze the /f_0sxfv1exQCiSZErKsV5.transcript . 2. Clearly identify every segment of continuous rep speech. Continuous rep speech is broken only by prospect interruptions lasting 10 seconds or more. 3. For every prospect interruption that occurs between rep speech segments, explicitly identify: Interruption timestamp (start and end) Exact interruption duration in seconds 4. Calculate whether the interruption breaks (≥10 seconds) or does not break (<10 seconds) the monologue. If the monologue was ≥10 seconds, treat that as a break and start a new rep segment after the prospect finishes talking. If the monologue was not broken, ignore the interruption and continue counting the rep's monologue. 5. Repeat the process until the end of the call. 6. Identify and return the rep's longest monologue, including its duration in seconds. 7. Output only the monologue's duration in the specified format. #EXAMPLES# Input: Transcript with alternating rep and prospect turns, with some prospect turns under and some over 10 seconds. Expected Output: DurationSeconds: [120]
easy, tell prospects to never cough on a call 🤣
Hey, That's a tricky problem with analyzing speech patterns. A few ideas that might help improve accuracy: Try a different approach with the prompt: ** Break it into steps: first identify all speaker segments with timestamps and extract them with Formula, then calculate durations, then apply the 10-second rule ** Ask the AI to output the analysis in a structured format showing each segment before giving the final answer ** Use a formula to pre-process the transcript and mark clear speaker transitions Alternative ** Try processing the transcript in chunks rather than all at once - sometimes AI handles smaller sections more accurately ** Consider using multiple AI columns to cross-check results *Prompt refinement: Your current prompt is solid, but you might try asking it to "think step by step" and show its work, or use a two-step process where it first maps out all segments, then calculates the longest monologue.
Have you tried testing it on a known transcript where you manually verified the longest monologue? Let me know if you have more questions.
This thread was picked up by our in-app web widget and will no longer sync to Slack. If you are the original poster, you can continue this conversation by logging into https://app.clay.com and clicking "Support" in the sidebar. If you're not the original poster and require help from support, please post in 02 Support.