• Added Keyword Spotting technology (high-level description) in REST API and GUI
• Added Age Estimation technology (high-level description) in REST API and GUI
• The Authenticity Verification technology has been improved with
  • further improved Deepfake Detection model 2.2.0 with significantly reduced number of false positive detections
  • new Deepfake Detection scores output as Log Likelihood Ratio (LLR)
  • added Audio Manipulation Detection subtechnology (high-level description)
  • added Replay Attach Detection subtechnology (high-level description)
• Further improvements of Virtual Appliance documentation:
  • more details about HW and SW requirements
  • more detailed deployment instructions for commonly used hypervisors

This release includes internal changes in licensing.
These changes are backwards compatible and should not affect existing users, with one exception - existing users of older Deepfake Detection model need to update to the new model and license to continue using Deepfake Detection in 4.0.0 release.

REST API changes

• Increased API default file size limit to 500 MB (limit can be changed manually in configuration file, see documentation)
• Added validation of voiceprint size in the REST API to prevent potential abuse via unusually large payloads masquerading as valid voiceprints
• Introduced a new Location header; the previous X-Location header is now deprecated
• Fixed missing X-Location header in some REST API responses (bug introduced in VA 3.7.0)
• Fixed an issue where the GET /api/task/:task_id endpoint sometimes returned HTTP 404 even though the requested task existed
• Various API documentation fixes and improvements, e.g. some query parameters minimum values are now documented correctly

GUI (web application) changes

• Increased GUI default file size limit to 100 MB (Limit can be changed manually in configuration file, see documentation)
• Fixed score format in Gender Identification exports: score is now exported correctly as percentage, not as decimal number
• Minor changes in homepage tile order and style
• Minor visual changes in the file table component (all technologies except Speaker Identification)

• Gender Identification technology (high-level description) is now fully working in the GUI.
• Emotion Recognition technology (high-level description) is now fully working in the GUI.
• New improved model 2.0.0 for Deepfake Detection subtechnology (high-level description).
• Virtual Appliance documentation improvements for easier navigation and better comprehension
• Configuration and administration changes:
  • The automation after uploading a package via Filebrowser GUI has been extended: after extracting the uploaded ZIP, it also runs the (auto)configuration script.
  • The automatic extraction now recognizes any ZIP file starting with "licensed-models" in the filename (i.e. also filenames like licensed-models_foobar.zip).
  • Fixed an issue where the automatic extraction could take a long time because it was performed multiple times.
  • When Virtual Appliance is deployed to Amazon EC2, the root and SSH password login is automatically disabled
  • Added more information to the output of the run-diag-report.sh diagnostic script
  • Removed the obsolete configuration script

Known issues:

• The root and SSH password login is disabled when custom cloud-init is used. This will be fixed in the next release.

• Added Authenticity Verification technology with Deepfake Detection subtechnology (high-level description) in REST API and GUI.
• Added Gender Identification technology (high-level description) in REST API (still only preview in GUI). The endpoints are differentiated based on the input type, which can be either a media file or a list of voiceprints from Voiceprint Extraction.
• Added Denoiser technology (high-level description) preview in GUI.
• Added Emotion Recognition technology (high-level description) preview in GUI.
• Updated Enhanced Speech to Text Built on Whisper model (high-level description) with a query parameter for word-level segmentation that also improves the overall accuracy of the timestamps. Because this behaviour is resource heavy, it is turned-off by default.
• Added settings for parallel threads and multiple instances for GPU support for Language Identification and Voice Activity Detection.
• Configuration and administration changes:
  • The Virtual Appliance startup process now displays system messages again for more clarity.
  • Improved detection of "system is ready" state during startup process
  • When licensed-models.zip package is uploaded via the Filebrowser GUI, it's automatically unpacked after upload.
  • New script configure-speech-platform.sh for the Speech Platform configuration, with more functionality. Use configure-speech-platform.sh --auto-configure to automatically configure the system according to models and licenses uploaded to /data/ folder. The enable-technologies.sh script is now obsolete and will be removed in next release.
  • Configuration YAML file is now much shorter, simpler and more comprehensive.
  • Turning on GPU support in configuration file is now easier, all GPU images are now included, it's not needed to download/configure them separately.

• Added consumption counting and GUI capacities indicator.
• Configuration and administration changes:
  • Virtual Appliance system console now displays the prompt only after the internal system is started. The system console is kept blank during that time. It may take some additional time for the GUI to become fully ready.
  • All technologies are now disabled by default after first start.
  • Added enable-technologies.sh script for enabling technologies according to uploaded models and licenses.
  • It's now possible to change UI limits for all technologies in the configuration file.

New available models with improved Voice Activity Detection configuration:

• Enhanced Speech to Text Built on Whisper: model 1.1.0
• Speaker Identification: model 5.2.0
• Language Identification: model 5.3.0
• Speaker Diarization: model 5.1.0
• Voice Activity Detection: model 5.3.0

• Added Audio Quality Estimation technology (high-level description) in REST API.
• Voice Activity Detection technology (high-level description) is available in both REST API and GUI.
• Speech Translation technology is now fully working in GUI.
• Speaker Diarization technology (high-level description) is now fully working in GUI.
• Updated Speech to Text Phonexia with the ability to use Preferred phrases.
• Updated the Speaker Identification model to xl-5.1.0, which is capable of carrying out automatic adaptation to various input audio sources (YouTube, Skype, WhatsApp, VoLTE, AMBE).

• Added Speech Translation technology preview in GUI.
• Added Speaker Diarization technology in (high-level description) REST API (still only preview in GUI).
• Added option to speed up Enhanced Speech to Text Built on Whisper via beamSize parameter in Virtual Appliance configuration file - smaller beamSize means faster processing (up to ~30% with large_v2 model and beamSize=1) at the expense of slightly lower accuracy.
• Speech to Text Phonexia and Time Analysis of Speech technologies updated to version 3.62.0.
• Configuration and administration changes:
  • Added support for importing Virtual Appliance to Microsoft Hyper-V.
  • Both system and data disk now automatically resize according to size set in virtualization software.
  • Customers can now use cloud-init with Virtual Appliance.
  • Added diagnostic script for collecting logs for troubleshooting.

• Language Identification technology now fully working in GUI.
• Added Gender Identification technology (high-level description) preview in GUI.
• Added possibility to share a GPU among multiple technologies, to better utilize the hardware resources.

• New and significantly easier way to install or update technology licenses - they are now loaded from a separate YAML file. Just unzip the licensed models package to the data disk and the system loads the licenses automatically.

  ⚠️ NOTE: The new license format is not compatible with previous versions, i.e. licenses used in earlier Virtual Appliance versions cannot be used in this version and a new license needs to be obtained from Phonexia.

• Added Language Identification technology preview in GUI.
• Added Speaker Diarization technology ([high-level description(/products/speech-platform-4/technologies/speaker-diarization)) preview in GUI.
• Added possibility to run Voiceprint Extraction on GPU to significantly boost the extraction performance.
• Virtual Appliance configuration changes are now automatically applied to corresponding components (no restarts needed anymore).
• Lower memory consumption during processing due to internal components optimizations.

• The Speech to Text Whisper Enhanced technology has been renamed to Enhanced Speech to Text Built on Whisper and a new Language switching feature was added. This feature identifies the predominant language spoken within each thirty-second interval of audio and the identified language is then utilized for transcribing that particular section.

• For Speech to Text Phonexia and Time Analysis of Speech technologies it's now possible to configure the number of tasks to be processed in parallel. It is done using the paralelism parameter in the corresponding sections of Virtual Appliance configuration file.

• Maintenance release with only configuration and administration related changes:
  • Models for Speech to Text Phonexia and Time Analysis of Speech technologies. are now loaded from data disk, not from image.
  • Speech to Text Phonexia and Time Analysis of Speech technologies updated to version 3.61.0.
  • Added extra environment variables for Speech to Text Whisper Enhanced.
  • Added maximum upload file size specification for filebrowser.
  • Moved Prometheus storage to data disk.