New

GPU & CPU

• Word Replacement - Enables words in the transcript to be modified using a search and replace pattern. Refer to documentation here for details.

• Supports the prefer_current_speaker configurable parameter to reduce the likelihood of incorrectly switching between similar sounding speakers. Refer to documentation for more details.

• Supports End of Utterance feature. When enabled, this helps detect the end of turn in a conversation. This can benefit use cases such as Voice Agents, dictation and translation, reducing latency. Refer to the documentation for more details.

• Supports speaker_sensitivity parameter to configure the sensitivity of speaker detection. Refer to documentation for more details.

• New Tagalog language is now available. Supports code-switching between Filipino and English for bilingual speech

GPU

• Channel diarization – Enables perfect speaker separation when there is one speaker per channel. Refer to documentation here for details

• Channel and Speaker diarization – Enables separation of multiple speakers per channel. Refer to documentation here for details

• Force end of utterance – Enables the client to force finalise transcription at the end of speech for faster finals (200ms), ideal when using external VAD or turn detection models for voice agents. Refer to documentation here for details

• Standard and Enhanced operating point

  • New bilingual transcription languages: Malay English (en_ms), Tamil English (en_ta), Mandarin English (cmn_en), Arabic English (ar_en). Refer to documentation here for details.

  • New Multilingual transcription language pack Mandarin Malay Tamil English (cmn_en_ms_ta) available now. Refer to the documentation for more details.

• Enhanced operating point

  • New medical domain-specific models for Danish, Dutch, English, Finnish, French, German, Norwegian, Spanish and Swedish giving the highest accuracy for healthcare use cases. Refer to the documentation for more details.

Improvements

GPU & CPU

• Speaker Diarization – Improved speaker change detection accuracy for long audio streams (1+ hours)

GPU

• Significant increase in session density for GPU inference.

• Standard Operating Point

  • Enables faster transcription and higher throughput, refer to the documentation for more details

  • New models released with notable accuracy uplifts for the below languages:

    Domain	Language	Relative Improvement
    General	Belarusian	68.9%
    General	Bulgarian	13.1%
    General	Catalan	57.5%
    General	Croatian	5.0%
    General	Czech	3.4%
    General	Danish	34.3%
    General	Esperanto	37.9%
    General	Estonian	5.1%
    General	Persian	9.7%
    General	Finnish	38.8%
    General	French	3.1%
    General	Galician	85.4%
    General	Greek	7.2%
    General	Hebrew	2.1%
    General	Hindi	9.3%
    General	Hungarian	15.8%
    General	Indonesian	19.1%
    General	Japanese	2.1%
    General	Korean	17.1%
    General	Latvian	25.1%
    General	Lithuanian	11.2%
    General	Malay	4.2%
    General	Marathi	28.8%
    General	Mongolian	19.6%
    General	Norwegian	18.5%
    General	Polish	2.0%
    General	Romanian	8.5%
    General	Slovak	4.1%
    General	Slovenian	9.8%
    General	Swedish	5.4%
    General	Thai	22.1%
    General	Turkish	33.2%
    General	Ukrainian	29.9%
    General	Urdu	16.1%
    General	Vietnamese	110.9%
    General	Welsh	20.3%

• Enhanced Operating Point

  • New models for English with improved accuracy in transcribing initialisms

  • New models released with notable accuracy uplifts for the below languages:

    Domain	Language	Relative Improvement
    General	Danish	20.6%
    General	Dutch	16.0%
    General	Spanish	4.1%
    General	Finnish	31.0%
    General	French	2.1%
    General	Interlingua	8.7%
    General	Japanese	5.8%
    General	Malay	6.8%
    General	Maltese	5.5%
    General	Norwegian	9.3%
    General	Swedish	10.4%
    General	Uyghur	8.0%

CPU

• Standard Operating Point

  • New models released with notable accuracy uplifts for the below languages:

    Domain	Language	Relative Improvement
    General	Belarusian	61.3%
    General	Bosnian	12.2%
    General	Bulgarian	6.4%
    General	Catalan	19.9%
    General	Welsh	5.4%
    General	Danish	2.8%
    General	Esperanto	43.9%
    General	Greek	4.1%
    General	Persian	5.9%
    General	Finnish	40.9%
    General	Galician	64.3%
    General	Irish	3.1%
    General	Hindi	4.0%
    General	Interlingua	24.9%
    General	Korean	7.6%
    General	Latvian	4.2%
    General	Malay	2.9%
    General	Maltese	3.5%
    General	Mongolian	19.4%
    General	Marathi	11.1%
    General	Norwegian	4.1%
    General	Romanian	2.7%
    General	Swedish	5.3%
    General	Thai	10.1%
    General	Ukrainian	19.3%
    General	Vietnamese	15.1%

• Enhanced Operating Point

  • Improved Speaker Diarization accuracy for Enhanced Operating Point

  • New models released with notable accuracy uplifts for the below languages:

  Domain	Language	Relative Improvement
  General	Belarusian	61.7%
  General	Bosnian	8.7%
  General	Bulgarian	7.4%
  General	Catalan	6.2%
  General	Welsh	3.3%
  General	Danish	27.4%
  General	Greek	3.2%
  General	Esperanto	41.8%
  General	Persian	4.7%
  General	Finnish	55.0%
  General	French	4.5%
  General	Irish	5.7%
  General	Galician	72.8%
  General	Hindi	3.1%
  General	Interlingua	35.1%
  General	Korean	4.1%
  General	Latvian	3.1%
  General	Mongolian	21.4%
  General	Marathi	12.4%
  General	Dutch	12.9%
  General	Norwegian	11.7%
  General	Swedish	5.9%
  General	Thai	7.0%
  General	Ukrainian	18.7%
  General	Uyghur	2.0%
  General	Vietnamese	20.1%

Fixes

• Fix failure to process some audio files starting with non-speech audio when speaker diarization is enabled.

• Fix for failure to generate Final transcripts for numbers when punctuation is disabled.

• Fix for Japanese to address decimals being occasionally transcribed as full stops.

• Fix for number formatting in Malay & English bilingual (en_ms) and Tamil & English bilingual (en_ta) language packs

  • Resolves the following tickets: 30055

• Fix for incorrect casing in Japanese (ja) transcription output

  • Resolves the following ticket: 28276

• Fixed a session failure when a custom dictionary’s first item is only a hyphen.

Security fixes

Vulnerability Management

• A Software Bill of Materials (SBOM) is available for download from the corresponding release page in our Support Portal.

• libgstreamer (CVE-2025-3887): This component has been manually patched to address CVE-2025-3887. Note: Security scanners may still flag this component as vulnerable because the base version string remains unchanged

Non-Applicable CVEs

The following vulnerabilities were reviewed and determined to have no security impact on this release due to specific configurations or the use of closed, trusted environments:

New

• GPU & CPU

  • New Virtual Appliance architecture with support for our latest generation of Ursa GPU models

  • New languages - Hebrew (he), Persian (fa), Irish (ga), Maltese (mt), Urdu (ur), Bengali (bn) and Swahili (sw)

  • Automatic Usage Reporting is enabled by default

  • Audio Filtering: pre-process audio to remove low-volume background speech which might otherwise be detected and transcribed. Refer to documentation here to get started

  • Disfluency removal: automatically remove disfluencies from your transcript. Refer to documentation here to get started

• GPU

  • GPU Ursa models - all 58 languages are now available on GPU

    • Major transcription accuracy gains

    • Major improvement in Speaker Diarization accuracy

    • Faster transcription

  • Bilingual Spanish and English language pack - this enables Spanish and English to be transcribed accurately within the same file

  • Audio Events: Detection of music, laughter and applause in media files now supported. Refer to documentation here to get started

Removed

The legacy Speaker Change Detection feature is now obsolete. Any sessions using the speaker_change and channel_and_speaker_change parameters will be rejected

Improvements

• GPU

  • Ursa2 models released, giving a broad accuracy uplift across languages (Enhanced Operating Point only):

    • Uplift for all languages, including a major improvement for Arabic dialects

    • Updated vocabulary for English

  • Improved music detection accuracy in Audio Events

• GPU & CPU

  • Ursa2 models released, giving a broad accuracy uplift for below languages (Standard operating point only):

    • Uplift for Basque (eu), Estonian (et), Polish (pl), Swedish (sv), Tamil (ta), Turkish (tr), Uyghur (ug)

  • Improvements to speaker diarization accuracy

  • Lower-latency Finals: the minimum allowed value of the max_delay parameter has been reduced from 2 to 0.7, enabling lower-latency final transcripts. Refer to documentation here for more details

  • Faster real-time transcription with more consistent latency

  • Major efficiency improvements for real-time transcription on GPU for English, French, German and Spanish with the Enhanced operating point

  • Partial transcripts now include numeral formatting, enhanced punctuation, and better casing.

  • Improved accuracy for recognition of Arabic numbers and currency

Fixes

• Written form for negative percentages in German transcription is now output as "%" instead of "Prozent"

• "Kyiv" output consistently in English transcription

• Updated English profanity tagging to remove a small number of non-profane words

• Fixed an issue with Mandarin (cmn) where consecutive English words were concatenated together

• Fix for delay in emitting Final transcripts for a subset of languages when using config containing punctuation_overrides={"permitted_marks":[],...

• When the server receives an EndOfStream message, all AddAudio messages received are dropped, and on the first AddAudio received after EndOfStream, the server will send an error message to the client (but not close the connection)

• Security fixes. A Software Bill of Materials (SBOM) is available for download from the corresponding release page in our Support Portal.

Known Issues

Security

• Vulnerabilities in Nvidia gpu-operator. We upgraded to the latest version, but some vulnerabilities remain unresolved by the vendor. These packages are inaccessible from outside the appliance.

• Vulnerabilities in packages shipped with k3s version v1.29.14+k3s1. We upgraded to the latest version, but some vulnerabilities remain unresolved by the vendor. These packages are inaccessible from outside the appliance.

• Vulnerabilities in version 12.0.1 of our Real-Time Container; see the 12.0.1 Container release notes for details.

New

The JSON-v2 output version is now 2.9

Improvements

• Language vocabulary improvements for French (fr), Italian (it), Hindi (hi), and Korean (ko)
• Remodelled German (de) language pack to utilize subwords, separating words into smaller segments to reduce Word Error Rate (WER)
• Improved numeral formatting in English
• Improved formatting for common telephone numbers, measurements, websites, email addresses and credit cards
• Alphanumerics now have upper-case letters
• Added regional handling for en-AU and en-US output locale to keep 'pounds' as words
• A number of other improvements and fixes for better readability

Fixes

• Fix for missing accented characters in Dutch transcription
• Security fixes