I’m currently tuning a speech-enabled IVR flow for a pharmaceutical client. We’re using the ‘Collect Input’ action with ASR (Automatic Speech Recognition) to allow customers to say the name of the medication they need refilled.
The issue I’m seeing is that for complex multi-syllabic drug names (like ‘Atorvastatin’ or ‘Levothyroxine’), the engine frequently misinterprets the input, even with a high confidence threshold. What’s worse, the ‘No Match’ path isn’t always triggering because the engine is confident it heard something else entirely.
I’ve added several of these names as ‘Synonyms’ in the grammar list, but the accuracy is still hovering around 65%. Has anyone successfully used ‘Speech Styles’ or custom grammars to improve recognition for niche technical terms like this?
The built-in ASR grammar is a generic model. For highly specialized vocab like pharma, you’re going to hit a wall with standard ‘Synonyms’.
One trick I’ve used on our healthcare dashboards is to switch to a ‘Closed’ grammar if possible. If the list of medications is finite (even if it’s 500 items), providing a static list in the grammar can significantly boost accuracy because the engine isn’t trying to match against the entire English dictionary. Also, check your ‘Speech Sensitivity’ setting—if it’s too high, it’ll pick up background noise as part of the word and throw off the phonetic match.
From a trunk perspective, also ensure you aren’t seeing any jitter on the BYOC side. ASR is extremely sensitive to packet loss. If even 1% of the audio packets are dropped during that multi-syllabic word, the phonetic engine will fail. I’ve seen cases where ‘jitter’ was misidentified as ‘ASR failure’.
We recently tuned our US-West trunks for a similar medical flow and found that reducing the ‘Inter-Digit Timeout’ helped slightly, though that’s more for DTMF. For ASR, look into using the ‘Confirmation’ logic in Architect—ask the customer ‘Did you say Atorvastatin?’ before proceeding.
To build on the confirmation logic: in my executive reporting, I track ‘ASR Success Rate’ as a KPI. When we implemented the ‘Did you say X?’ confirmation, our ‘Self-Service Completion’ rate jumped by 20%. It adds a few seconds to the call, but it’s much better than a frustrated customer being routed to an agent because the bot didn’t understand ‘Levothyroxine’ for the third time.