STT Release Notes

The following sections list every released version of STT, newest first. Each section comes from the monthly OCP® Release Notes page named in its Released line.

STT 8.12.0

Released: October 2026

• This release contains internal technical and infrastructure work. There are no customer-facing changes in this release.

STT 8.11.0

Released: October 2026

Improvements

• Improved Resilience for deepVB® Verification — deepVB® verification now responds faster and no longer blocks call processing when the verification service experiences delays, reducing the chance of calls stalling during authentication.

• More Reliable Speech Recognition Reranking — Resolved an intermittent issue in the recognition reranking step that could cause requests to fail, improving the consistency of speech recognition results.

Fixes

• Eliminated Call Hangs From Delayed Verification Billing — Calls using deepVB® verification no longer stall or go unanswered when the billing service responds slowly. Call setup and teardown now complete promptly, even during billing service delays.

• This release includes security updates.

STT 8.10.1 Hotfix

Released: September 2026

Fixes

• This release includes security updates.

STT 8.10.0

Released: September 2026

Improvements

• Improved Recognition Accuracy During Brief Pauses - deepASR® recognition results now capture only the audio spoken before a short pause, so words spoken right after a pause no longer get attributed to the wrong part of the conversation. This delivers more consistent, accurate results in conversations with natural short pauses.

STT 8.9.0 Release

Released: August 2026

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.

STT 8.8.2

Released: July 2026

Improvements

• More Complete Speech Transcripts - deepASR® now captures the full spoken sentence more consistently. Previously, a trailing word could occasionally be missing from the transcript under certain call conditions; this gives your applications a more accurate, complete record of what callers said.

STT 8.8.1

Released: July 2026

Improvements

• More Complete Speech Transcripts - deepASR® now captures the full spoken sentence more consistently. Previously, a trailing word could occasionally be missing from the transcript under certain call conditions; this gives your applications a more accurate, complete record of what callers said.

• More Accurate Zip Code Recognition - Fixed an issue where dictating a zip code slowly, digit by digit with pauses, could occasionally insert a stray extra letter in the middle of the number. Numbers dictated this way are now recognized more accurately.

STT 8.8.0

Released: July 2026

New Features

• Improved Liveness Detection Accuracy - deepASR® now streams a brief period of silence before speech audio begins, giving the liveness detection engine more complete audio context. This reduces misclassification at the start of calls, providing more reliable detection of whether a live person or an answering machine has picked up.

Improvements

• More Reliable Number Transcription for GENERIC ASR (logos models) - deepASR® now applies a post-processing step to Generic ASR transcription results, converting spoken numbers (zero through twelve, including "oh" for zero) into digits when the engine's built-in normalization does not. This reduces transcription errors in scenarios where callers speak numeric values such as zip codes, account numbers, or menu selections.

• Faster and More Consistent Recognition Completion - Resolved a condition where simultaneous internal processing events could cause recognition to stall for up to two seconds before returning results. Speech recognition now completes promptly and consistently, regardless of the order in which audio processing events arrive.

Fixes

• Reliable Session Log Uploads - Session log files were not being uploaded after early April 2026. This has been resolved by consolidating the upload service directly into deepASR®, ensuring session logs are captured and stored reliably for every call.

• Prevented Engine Failures on Large Audio Frames - deepASR® now enforces a maximum size on incoming audio chunks before they are forwarded to the recognition engine. Previously, oversized audio frames could corrupt the internal audio stream, causing the recognition engine to fail for the remainder of the session and forcing a session restart.

• Prevented Startup Crash When Voice Biometrics is Misconfigured - deepASR® no longer crashes on startup when the biometrics service is disabled but the voice verification service is also enabled. The service now detects the inconsistency and disables voice verification automatically, ensuring a clean startup in all configuration combinations.

• Resolved Stuck Engine Incident Causing Call Failures - A production incident where all deepASR® recognition engines became occupied and unable to process new calls has been investigated and resolved. Improvements to thread management and internal timeout handling prevent engines from becoming permanently stuck, restoring normal call processing without requiring a restart.

• This release includes security updates.

STT 8.7.2

Released: June 2026

Fixes

• In depth metrics for hybrid turn detection

• Performance improvement in step audio redaction feature

• This release includes security updates.

STT 8.7.0

Released: May 2026

New Features

• Hybrid Turn Detection - deepASR® now combines audio signals and partial speech transcriptions to determine when a caller has finished speaking, replacing the previous text-only approach. This delivers more natural conversation pacing, reducing instances where the system responds too early or waits unnecessarily before handing the turn back to the virtual agent.

• Multiple Address Assets per Recognition Step - Address recognition now accepts multiple postal-area asset sets in a single step, allowing the system to cover a broader range of addresses without requiring additional dialog turns. This reduces friction in address-collection flows and improves first-attempt recognition rates across complex or overlapping coverage areas.

• Enhanced Liveness Detection Model Configuration - The Liveness Detection capability now retrieves the active detection model and decision threshold directly from your group configuration at the start of each call, ensuring the correct anti-spoofing parameters are applied automatically without manual intervention.

Improvements

• More Natural End-of-Turn Recognition - Turn-taking decisions are now deferred until a brief silence is detected, preventing the system from interpreting natural mid-sentence pauses as the end of a caller's response. Callers experience fewer interruptions during extended or complex utterances.

• Accurate Word Timing with Omilia VAD - Word-level timestamps are now accurate when Omilia's built-in voice activity detection is active alongside the generic transcription model. This improves downstream features that rely on precise timing such as redaction, analytics, and turn detection, without the higher processing cost of the previous workaround.

• Automatic Audio Format Correction - Incoming audio that does not match the expected sample rate is now detected and resampled automatically before transcription. This safeguards recognition quality when audio arrives in an unexpected format, preventing silent accuracy degradation.

Fixes

• This release includes security updates.

STT 8.6.2

Released: May 2026

Improvements

• Reliable Final Results with Legacy Voice Detection - Speech recognition now correctly handles empty responses when using legacy voice activity detection with small audio inputs, eliminating infinite waits and ensuring timely no-input results. This resolves an issue where short audio segments would cause the system to hang indefinitely waiting for final transcription results.

STT 8.6.1

Released: May 2026

Fixes

• Accurate Logging for Dynamic Grammar Recognitions - Speech recognition logs now correctly reflect when dynamic grammars are used. Previously, successful dynamic grammar calls were incorrectly logged with a generic placeholder value, disrupting analytics and reporting. The system now accurately captures which grammar type handled each request, restoring reliable metrics for recognition analysis.

STT 8.6.0

Released: May 2026

New Features

• Word Boosting for Generic ASR Models - Generic ASR models (ATLAS and HERMES) now support on-the-fly word boosting, allowing you to increase recognition accuracy for specific terms during speech recognition. For more information, go to ASR Adaptation User Guide.

Improvements

• Enhanced Numeric Transcription Formatting - Generic ASR models now returns numeric and alphanumeric sequences without separators (for example, 1231234123 instead of 123-1234-123), significantly improving PCI and PII term detection accuracy.

• Turn Detection Completion Reasons - The Riva integration now accurately labels turn completion reasons, distinguishing between maximum speech duration (MX) and text-based turn completion (TC) instead of marking all results as speech recognition completion (SR).

• Word-Level Confidence Scores - Word confidence scores from the Riva integration now return valid probability values between 0.0 and 1.0 instead of placeholder negative values, enabling proper low-confidence threshold logic.

• Delayed Audio Upload Handling - The delayed audio upload feature now handles client requests gracefully even when the feature is disabled on the ASR node, preventing unexpected 501 errors.

Fixes

• Word Timings Accuracy - Fixed word timing drift that caused increasing discrepancies over the duration of calls, ensuring consistent timing accuracy for audio redaction and analysis.

STT 8.5.0

Released: March 2026

What's new

• Sensitive Term Audio Redaction - Introduced a new pipeline step for automatic audio redaction of sensitive terms, enhancing privacy and compliance capabilities.

Fixed

• Speech End Handling - Improved generic ASR speech end detection reliability, along with code refactors to support future extensibility.

• Default Grammar Loading - Fixes and improvements to default grammar loading behavior for more consistent and predictable ASR initialization.

• Logging - Enhanced logging across the service for improved observability and debugging.

• Internal API - Various internal API improvements to support stability and integration reliability.

STT 8.4.2

Released: February 2026

Fix for faulty persistent value of increment_step_audios_on_storage flag.

STT 8.4.1

Released: February 2026

Hotfix for ASR Generic transcription.

STT 8.4.0

Released: February 2026

Features

• Configurable variable for GENERIC ASR. (US1 only)

• Additional models with linked URLs to be supported as per multilingual one.

Improvements

• Identify speech frames on GENERIC ASR integration for VB. (US1 only)

• Correct return response for Native cleanup request.

• Verify ASR Generic speech-end detection with VAD.

Fixes

• Ensure that deepASR logs are valid ECS schema JSONs.

• Investigate reason for failing CAPI Turn-detection test cases in VFO-Prod.

• Fixed stuck engines on deepASR.

• VFO: Turn Detection returns SR instead of TC despite identical configuration.

• Turn-Taking Delay Issues in VFO-Prod.

• DeepASR tries to parse the audio name as dialogID.

• DeepASR ignores the defaultGrammar option in VFO-prod.

• Investigate agent assist stuck engines.

Security

• Resolved vulnerabilities

STT v8.1.1

Released: October 2025

Fixes & Improvements

• Allow setting the step number on Stream WS

STT v8.1.0

Released: October 2025

New Features

• Implement OTF params for nBest and latNBest

• Automatically calculate the RTP port range based on the licenses/activeEngines number

• Mask sensitive logs (e.g., secret billing tokens)

• Remove MRS isolation mode to simplify behaviour and solve issues with autoscaling

• deepASRCore: further cc formatting improvements

• Code quality fixes suggested by Cursor

• deepASRtools: Make the dithering parameter configurable in the omDecoder and release deepASRTools 3.3.0

Fixes & Improvements

• Fix invalid model_fst object crash during recognition

STT v7.13.1

Released: April 2025

Fixed

• Fix for the ASR tags that have been annotated

STT v7.12.0

Released: March 2025

• Remove deprecated keycloak libraries

• Security Fixes