Skip to main content
Automatic Speech Recognition (ASR) determines what the agent hears. Every downstream system depends on this layer: Knowledge Base retrieval, routing, response control, and handoff logic. Small transcription errors here can silently break otherwise well-designed agents. Global ASR configuration is where you correct systematic listening problems before compensating elsewhere.

​
What Global ASR is for

Use Global ASR to:
  • Reduce unnecessary clarification questions.
  • Improve routing accuracy for calls.
  • Stabilise KB selection for spoken queries.
  • Correct recurring transcription errors at the source.
It is not a substitute for good KB design or clear agent behaviour rules. Think of it as sharpening the agent’s ears, not teaching it new facts.

​
Two tools you should understand

​
1. Keyphrase Boosting

Keyphrase Boosting biases the ASR model toward recognising specific words or phrases more reliably. Use this when:
  • A word is important for routing or actions.
  • Misrecognition causes the agent to take the wrong path.
  • Users consistently repeat themselves for the same term.
Typical candidates:
  • Department names
    “front desk” “billing” “room reservations”
  • Product or service names
    “bell desk” “late checkout”
  • Action-critical verbs
    “cancel” “confirm”
How to configure:
  • Add the keyphrase exactly as you want it recognised.
  • Start with Default bias strength.
  • Increase toward Maximum only if errors persist.
Guidance:
  • Bias strength is cumulative. Over-biasing many terms can degrade overall accuracy.
  • Avoid boosting generic words like “room” unless they are routing-critical.
  • Prefer multi-word phrases over single words when possible.
Example If callers asking for the “bell desk” are frequently routed incorrectly, boost “bell desk” rather than boosting “desk”.

​
2. Transcript Corrections

Transcript Corrections rewrite text after transcription, mapping common errors to the intended phrase. Use this when:
  • ASR consistently hears the same wrong thing.
  • Accents or homophones cause predictable errors.
  • Brand or proper names are repeatedly mangled.
Typical use cases:
  • Homophones
    “copper” → “hopper”
  • Accent-driven substitutions
    “reef fund” → “refund”
  • Brand names
    “Hopper Consumer” → “Hopper”
How to configure:
  • Choose a match type (substring or regex).
  • Define the misheard expression.
  • Define the corrected replacement text.
Guidance:
  • Prefer narrow matches over broad regex.
  • Avoid corrections that could unintentionally rewrite unrelated phrases.
  • If unsure, test in Sandbox with real call audio before promoting.
Example If “cancel” is sometimes transcribed as “council,” add a correction rather than increasing bias on “cancel” alone.

​
A practical workflow

  1. Review recent calls where:
    • The agent asked for repetition.
    • The wrong KB topic was selected.
    • Routing or handoff failed unexpectedly.
  2. Identify the specific word or phrase that failed.
  3. Decide:
    • Boost if the word is correct but under-recognised.
    • Correct if the word is consistently misheard as something else.
  4. Apply the smallest possible change.
  5. Test again in Call.
  6. Promote only after confirming improvement.

​
Common mistakes

  • Boosting everything “just in case”.
  • Using Transcript Corrections to fix KB design problems.
  • Applying maximum bias globally without testing.
  • Forgetting that local (flow-level) ASR settings override global ones.

​
Verify

You’re done when:
  • Previously misheard phrases are transcribed correctly.
  • The agent asks fewer clarification questions for known terms.
  • Routing and KB selection stabilise in live call testing.
  • Behaviour remains consistent after promotion between environments.
Global ASR tuning is iterative. Small, deliberate adjustments here pay off more than large rewrites elsewhere.