Summary: We are updating the Talkdesk CXA Operations Center™ (formerly AI Trainer) Custom Vocabulary interface to remove deprecated configuration fields ("Sounds Like", "IPA") and introducing a simplified workflow based on formatting and term weighting. This new approach aligns with the latest industry best practices from Speech-to-Text to improve model performance and usability.
Impacted customers: All customers using Talkdesk CXA Operations Center (formerly AI Trainer) for Speech-to-Text (STT) model optimization.
What is changing and why
Our speech providers have shifted away from manual phonetic inputs. These fields are no longer effective and often lead to confusion or misconfiguration.
We are removing the legacy "Sounds Like" and "International Phonetic Alphabet (IPA)" fields from the Custom Vocabulary creation flow. Below, we’ve highlighted the main changes for reference:
- UI Update: "Sounds Like" and "IPA" columns (CXA Operations Center > Custom Vocabulary) are being removed from the UI. “Wrong Transcript” column will be renamed to “Phrase”.
- New Workflow (Weighting): The logic has shifted to a weighting mechanism that adds a word to increase its importance, making it more likely to be recognized in transcriptions.
Impact
- Admins & Supervisors: Anyone currently improving STT models inside CXA Operations Center.
- Existing Models: Existing vocabulary entries will remain active. However, users may need to audit their lists. Not all vocabularies need changes - only those where the user previously relied on mapping a "wrong transcription" to a "right display" as a correction rule. These should be updated to follow the new "Custom Vocabulary" guidelines.
Action required
- Audit Recommended: Existing custom vocabulary will remain active. However, customers are encouraged to generally review and update their lists to ensure accuracy and relevance.
- Certain words or phrases can gain increased weight in speech-to-text (STT) transcriptions when mapped to a custom vocabulary entry.
- Please consult the Training Speech-to-Text Models article for more details.
Time and Duration
- Rollout Date: December 18, 2025.
- Type: One-time release to all regions.
Resources
- Training Speech-to-Text Models in Al Trainer
- Release Notes: December 18, 2025
Stay updated with important Talkdesk product alerts by subscribing to the Talkdesk Weekly Digest. See how here.
If you have any questions or concerns, please contact your Customer Success Manager or the Customer Care team via the Talkdesk Support Portal.