Skip to main content
Version: 2026.03.0

Deepfake Detection Changelog

4.1.0 (2026-03-03)​

Added​

  • [PLAT-1547] Update detection algorithm. The microservice supports the models with compatibility version 0 and 1 now.
  • [PLAT-1538] technology_name field to phonexia.grpc.common.Status/GetStatus gRPC method response
  • [PLAT-1544] Docker healthcheck using grpc-health-probe for container orchestration

Changed​

  • [PLAT-1539] Changed resampler library

Fixed​

  • [PLAT-1554] Microservice failing on certain audios
  • [PLAT-1538] Removed unused secure option from phonexia.grpc.common.Status/GetStatus gRPC method

4.0.1 (2026-01-13)​

Fixed​

  • [PLAT-1506] Fixed 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).
  • [PLAT-1509] Fixed GPU processing failure on NVIDIA Blackwell architecture GPUs (e.g., RTX 1000 series) caused by a bug in ONNX Runtime.
  • Fixed license validation error handling for legacy (v1) license files with missing required fields. Previously, this could cause undefined behavior.

4.0.0 (2025-11-24)​

Breaking change​

  • [PLAT-1424] - Improve compatibility with future models - Warning microservice requires at least model generic version 2.6.0

3.0.0 (2025-10-14)​

Changed​

  • Microservice now by default logs each request with incremental correlation id
  • [PLAT-1429] moved the gRPC API from phonexia.grpc.technologies.deepfake_detection.experimental to phonexia.grpc.technologies.deepfake_detection.v1
  • [PLAT-1471] Refactoring of the internal audio handling to reduce memory usage. Audio format support is now limited to WAV and FLAC.

Added​

  • [PLAT-1367] Support for a new model 'generic'. Please note that the model requires different license than model 'beta'.

Fixed​

  • [VOX-2666] Inconsistent logging
  • [PLAT-1277] Possible memory leak on cuda device

2.2.0 (2025-05-20)​

Changed​

  • [PLAT-1235] Output score is LLR (log-likelihood ratio) now
  • [PLAT-1235] Only voice segments in audio are processed to avoid false alarms
  • [PLAT-1235] The algorithm is more robust against random deviations in the score in a small part of the audio

Removed​

  • Compatibility with models <2.2.0

2.1.1 (2025-05-15)​

Fixed​

  • [PLAT-1243] Compatibility issue with older v2 models

2.1.0 (2025-04-29)​

Changed​

  • The microservice requires an authenticity verification license with deepfake detection enabled

Added​

  • [PLAT-1056] Support for new licenses tied only to model major version. Old licenses tied to precise model version are still supported.
  • [VOX-1293] phonexia.grpc.common.Status service
  • [VOX-1293] Field license_flags to LicensingInfoResult in phonexia.grpc.common.Licensing

2.0.0 (2025-04-02)​

Updated​

  • [PLAT-1185] Use model v2 with higher accuracy

Removed​

  • Compatibility with model v1

1.1.1 (2025-02-11)​

Fixed​

  • [PLAT-1134] Deepfake detection scores are invalid for some recordings

1.1.0 (2025-02-05)​

Added​

  • [PLAT-1070] Default port 8080 exposed in Dockerfile

Changed​

  • [PLAT-1107] Inverted deepfake score

1.0.0 (2025-01-08)​

Added​

  • Initial implementation of Deepfake detection service