Skip to main content

Version 2026.03.0

· 6 min read

General changes​

  • Versioning scheme changes from semantic (e.g., 5.0.0) to calendar in the format YYYY.MM.minor, e.g., 2026.03.0 corresponds to the first version in March 2026

Features​

  • Changes in configuration (speech-platform-values.yaml):
    • BREAKING CHANGE: The Speech to Text technology is now configured in the speech-to-text subchart, and the configuration has the same form as other microservices. Previously, it was configured in the speech-to-text-phonexia subchart and it used a grpcAdapter configuration for Speech to Text based on Phonexia Speech Engine
    • Technologies can configure instancesPerDevice to control how many tasks can be processed by a technology on a single CPU/GPU in parallel
    • Technologies can configure threadsPerInstance to control how many CPU threads will be used for each request
    • All on-demand technologies can set a cooldownPeriod to manage how long instances stay idle before they are scaled down
  • Virtual Appliance scripts no longer require root privileges and can be run under the default phonexia user
    • The scripts have been moved to /home/phonexia/scripts/
  • The maximum number of microservice instances running at the same time has increased to 300
  • The NVIDIA driver has been upgraded to 570-open, which enables using newer GPUs
  • After the Virtual Appliance starts, a warning message is now displayed on the console if the virtual machine has insufficient RAM or CPU

Fixes​

  • Large integers in configuration are now handled correctly
    • Values like file size limits and storage capacities were sometimes converted to scientific notation, causing deployment issues
  • The misleading "GPU only" description for instancesPerDevice in speech-platform-values.yaml is removed

Technology changes​

Features​

  • Deepfake Detection uses an updated detection algorithm that produces more accurate results
  • Pronunciation source in Keyword Spotting can be PARTLY_GENERATED (previously USER) if a multi-word keyword contains some generated and some built-in pronunciations
  • Phonexia Speech to Text uses a new internal implementation (previously, Phonexia Speech Engine was used). The API remains the same; only some error details have changed
    • As a result, users of the technology need to obtain models with an updated license when updating the virtual appliance to this version
  • Enhanced Speech to Text Built on Whisper, Keyword Spotting, Replay Attack Detection, and Voice Activity Detection use less memory to process audio
  • Many technologies that process audio use a new resampler library
    • As a consequence, these technologies can produce negligibly different results than the previous version

Fixes​

  • Many technologies that process audio include a fix for a rare issue where audio resampling could fail when processing large files, particularly when converting between certain sample rates (e.g., 11025 Hz to 8000 Hz)
  • Many technologies that process audio include a fix for a GPU processing failure on NVIDIA Blackwell GPUs (e.g., RTX 1000 series) caused by a bug in ONNX Runtime
  • Enhanced Speech to Text Built on Whisper
    • Fixes an issue where some segments did not contain word segmentation even if the feature was enabled
    • Translation now doesn't throw an unsupported language error when the language is supported for Translation but not for Transcription

REST API changes​

Features​

Fixes​

  • The healthcheck endpoint now recovers from an error caused by invalid technology licenses
  • Fixed a number of occasions where temporary files were not removed from storage

GUI (web application) changes​

Features​

Settings menu​
  • A new Licenses section shows a detailed listing of available licenses for technologies
    • Users can upload new license files and model archives (licensed-models.zip)
  • In the Microservices tab of the System configuration section, users can configure a cooldown period for on-demand microservice instances (see General Changes above)
Technologies​
  • Users can use the Send to... feature to send files (individually or in bulk) for processing by other technologies
  • All technologies use a unified file addition dialog
    • Users can select files already uploaded for other technologies
  • Files are sorted by upload date from the most recent to the oldest
    • Sorting can be changed with the Added Date column (hidden by default)
  • Gender Identification, Audio Quality Estimation, and Language Identification show top scores for individual channels in the quick results
  • Deepfake Detection now uses a new production-ready model
    • As a consequence, cached Deepfake Detection results are removed

Fixes​

Technologies​
  • Keyword Spotting now correctly shows phoneme keyboard for Arabic
  • Keyword Spotting shows error messages when entering invalid characters in spelling and pronunciation

Included Components
  • Age Estimation 2.1.0
  • Audio Manipulation Detection 2.1.0
  • Audio Quality Estimation 2.1.0
  • Deepfake Detection 4.1.0
  • Referential Deepfake Detection 2.1.0
  • Denoiser 2.1.0
  • Emotion Recognition 2.1.0
  • Enhanced Speech to Text Built on Whisper 2.2.0
  • Gender Identification 2.1.0
  • Keyword Spotting 2.2.0
  • Language Identification 2.1.0
  • Replay Attack Detection 2.2.0
  • Speaker Diarization 2.1.0
  • Speech to Text 2.2.0
  • Text Translation 2.1.0
  • Time Analysis 2.1.0
  • Transcription Normalization 2.1.0
  • Voice Activity Detection 2.2.0
  • Voiceprint Comparison 2.1.0
  • Voiceprint Extraction 2.1.0