August 2026

Agent VB Middleware 6.21.0

New Features

• Automatic PCI Data Protection During Voice Biometric Enrollment - OCP VB Agent-side Authentication® enrollment and verification now automatically detect and set aside any recorded speech containing sensitive payment card data, protecting customers without extra agent effort. Enrollment reaches a quick initial checkpoint sooner, while additional safeguards ensure only clean, nonsensitive speech counts toward enrollment. When sensitive content can't be confidently checked, that portion is skipped, so it's never retained.

Audio Hub 7.3.2

Fixes

• Correct Voice Used for Newest Omilia Voices - Requests for a specific voice from Omilia's newest voice family are now honoured. Previously, these requests were silently served with the deployment's default voice, so applications configured for one voice could be heard speaking in a different voice. Voice selection now behaves consistently across all supported voices.

• Uninterrupted Streaming After Quiet Periods - OCP Conversational Text-to-Speech® streaming no longer fails on the first request following a period of low traffic. The connection to the speech engine is kept warm, and a request that fails before any audio has been played is automatically retried, so callers hear the full prompt instead of an error.

Chat Connector 8.18.1

Fixes

• Reliable Sensitivity Flagging for Chat Input - Chat messages containing sensitive information, such as passwords or PINs, are now correctly flagged as sensitive so they're consistently masked in transcripts and logs. This resolves an issue where the flag wasn't always set as expected, strengthening privacy protection for customer conversations.

Security

• Updated Embedded Components - The embedded Apache Tomcat runtime (10.1.57) and the Netty networking library (4.1.136.Final) have been upgraded to address published security advisories. No configuration change is required.

Copilot 1.12.0

New Features

• Session pinning, renaming, deletion & search in Copilot Widget - Manage your chat history more easily: pin important conversations, rename sessions, remove ones you no longer need, and search across chat titles to find a conversation fast.

• Self-Learning Dashboard

Improvements

• More reliable chat history and onboarding journey.

• Smoother user management.

• Clearer tool configuration guidance - The tool configuration screen now highlights proven workflow patterns to help you set up agent tools faster.

• Better Knowledge Base uploads - FAQ documents are now grouped more clearly, and uploading a partial batch of FAQs is smoother.

• Improved persona setup - The persona editor now pre-fills detected persona details, reducing manual setup.

Copilot 1.13.0

New Features

• Fork Chat Sessions - You can now fork an existing Copilot chat into a new session. Explore a different path without losing the original conversation, and keep working from a clean branch of the same history.

• Richer Agent-Proven Workflow Review - Agent-proven recipes now show the conversation alongside each desktop step, and explain in plain language why a detected recipe is not yet ready to replay.

• Default Omilia Voice for New Self-Learning Agents - Newly created Self-Learning applications are configured with an Omilia Global TTS voice by default, so first voice tests sound polished without extra TTS setup.

• Simpler RingCX Connection - RingCX project setup now accepts an account ID, with a more reliable authorization flow and clearer success or error feedback in the connection popup.

Improvements

• Smoother Voice Test Waiting States - Pending indicators in Self-Learning voice mode fade in with light motion, so short waits feel less abrupt during testing.

• More Reliable Knowledge Base Answers - Concierge agents are less likely to refuse in-scope FAQ questions when the knowledge base covers a broad set of topics, so callers are routed to the right knowledge more consistently.

• Faster Call Ingestion - Call assessment during ingestion runs more efficiently under load, so new conversation data is ready sooner.

• More Reliable Large Knowledge Base Publishing - Large knowledge bases now publish in smaller steps with clearer progress during Self-Learning deploy, reducing timeouts on bigger projects.

Fixes

• Stronger Privacy in Call Assessment - Pre-redaction personal data is no longer retained in call assessment storage.

• Reliable Agent Assist Navigation - Opening Agent Assist now takes you to the correct destination again.

• Cleaner Agent Loading Experience - Agents that already have content no longer flash the first-time ingestion animation.

• More Stable Concurrent Updates - Background processing and UI edits are less likely to overwrite each other when they run at the same time.

• This release includes security updates.

Hybrid Turn Detection 2.1.0

New

Hybrid turn detection uses audio and text cues to label end-of-turn in customer–agent interactions. This service reduces average dialogue duration and improves the user experience by minimising wait time.

Liveness Detection 6.1.1

Improvements

• Change the logic so that the first 3 sec of an audio chunk, which is duplicated, are not counted in the overall average score.

• New calibration values for the model 3.1.0

OCP NLU 1.16.0

Improvements

• Larger File Uploads for Training and Evaluation - Training and evaluation file uploads now support files up to 15 MB, up from the previous 10 MB limit, giving you more room to work with bigger datasets.

• Clearer Snapshot List Layout - Fixed a display issue where the snapshot list could overlap the Cancel and Create buttons, so both stay visible and clickable.

• Non-Latin Character Support in Context Rules - Context rule matching patterns now accept non-Latin characters, such as Cyrillic script, directly from the console, matching what was already available through the API.

• Cleaner Model Import Error Handling - When linking a model to an existing artifact fails because of a domain mismatch, the system no longer leaves behind a stray application, keeping your workspace free of unwanted leftovers.

• Recover from Stuck App Checkouts - You can now unblock and delete an application that's stuck mid-checkout or mid-deletion because its underlying resources were already removed, so you're not left waiting on an unrecoverable state.

• Cross-Validation Results Travel With Your Model - Exporting, importing, or creating a snapshot of a model now automatically carries its cross-validation results along with it, so validation metrics move with the model instead of being left behind.

• OCP Console® Overhaul, Phase Two - The OCP Console® continues its refresh with updated modals, drawers, input fields, and data tables for a more consistent look and feel across every section. The new experience can be turned on or off through a feature flag, so you control when you switch over.

• Simplified Deployment Notifications - Notifications about inactive NLU deployments scheduled to stop are now grouped into a single message per user instead of one per application, with clearer wording letting you know deployments will stop in 24 hours.

• Clearer Limit Messages for Models and Apps - When you reach a model or application limit, the message now tells you specifically whether it's your account limit or your group's limit, and who to contact, so you know exactly how to proceed.

• More Reliable xPack Deployments - xPack deployments now check that required components are ready and configurations are in sync before proceeding, and prevent duplicate deployments of the same version, reducing failed or conflicting updates.

Orchestrator 1.36.2

Fixes

• Restored Volume and Rate Controls in Text-to-Speech Configuration - Some text-to-speech configuration screens stopped showing volume and rate settings for certain locales, so changes could not be saved. Volume and rate controls now display correctly every time, so you can configure them without interruption.

Pathfinder 1.9.4

Improvements

• Faster Response Times During High-Traffic Periods - Response times stay consistently fast during bursts of heavy concurrent usage, so authentication, search, and retrieval actions remain smooth even at peak load.

STT 8.9.0 Release

New Features

• Added --prepare-bgrammar CLI flag to omCompiler, enabling partners to generate .bgrammar files directly during compilation

Improvements

• More Reliable Number Transcription for GENERIC ASR (logos models) - deepASR® features an improved post-processing mechanism that converts spoken numbers (zero through twelve, including 'oh' for zero) into digits with higher accuracy.

• More Transcription for GENERIC ASR (logos models) - Improved GENERIC ASR (logos models) word capture to ensure complete final transcriptions.

Fixes

• Proper deletion of wav files - Fixed a bug where temporary .wav files were not deleted post-upload when delayed upload was enabled per-request while disabled globally.

• Fix duplicate final step - Fixed a race condition during audio stream shutdown that caused duplicate final transcription result reports to be emitted.

• Dynamic grammar compilation - Fixed a segmentation fault occurring during dynamic recipe compilation via the v2 endpoint.

TTS 3.2.0

Spanish pronunciation & verbalization fixes (es-US, es-CO)

• Chained spelled-out letters are now clearly separated.

• Capitalized abbreviations opening a sentence are handled correctly.

• Unit designators after address abbreviations are read properly as part of the address.

Colombian Spanish (es-CO) number & currency fixes

• Decimal values with units read correctly: "2,5Gb" → "dos coma cinco gigabits" (previously "veinticinco gigabits").

• Numbers written with a period are no longer mistaken for dates.

• Amounts under one unit read as the subunit: "$0,01" → "un centavo" (previously "signo de dólar cero coma cero uno").

• Decimal numbers using the Colombian "," separator are read as "coma" instead of "punto".

Lexis voices only: en-US, en-GB, fr-CA, es-US, es-CO

• Punctuation is no longer preceded by an audible gap.

• Long runs of the same repeated digit are correctly handled and do not result in hallucinated audio.

TTS 3.2.1

Stability (Lexis voices)

• GPU crashes observed under sustained high-concurrency traffic no longer occur.

• If a GPU-level error does happen, the service now reports itself as unavailable instead of continuing to accept traffic it cannot serve, so the affected instance is restarted automatically and recovers on its own. Previously it stayed in service, failed every request, and needed a manual restart.

• The hallucination guardrail is now on by default. Audio that runs far longer than the input text can justify is cut short and the prompt is retried. The retry can be preceded by a short apology (for example "Sorry, let me try again."), configurable per deployment.

Verbalization

• "Ave." is now read as "avenue", including when capitalized. The same fix covers other capitalized address abbreviations such as St. and Blvd.

• Numeric sequences are now read three digits at a time, regardless of whether they are repeated or not.

• Telephone numbers are read digit by digit in all supported locales, with two-zero groups spoken as multiples: "800" as "eight hundred", a trailing "1400" as "fourteen hundred".

Prosody

• Unexpected pauses in the middle of a sentence, most noticeable in Greek, are reduced. Where a long sentence still has to be broken up, the break is placed at a comma instead of in the middle of a clause, so the words either side of the break are no longer run together.

• This is a mitigation rather than a complete fix. The underlying cause, sentence length being estimated before the text is verbalized, requires a structural change to the engine and is planned for a following release.

xPacks 1.19.2

New OCP xPacks

• fr-BE Universal deepNLU® xPack v3.0.0