Changelog¶
das-Peak 2.21¶
Added¶
- Expanded our audio support, adding the
G711 a-law
codec for audio input. - New antispoofing engine including deep fake detection.
Fixed¶
- Error messaging. Now, if the service receives an audio with an invalid codec, it will respond with a more informative response message, helping to quickly identify and resolve the issue.
das-Peak 2.20¶
- Significant improvements in our spoof detection system.
das-Peak 2.19¶
- Deployment image efficiency enhanced: Container size diminished by 50%.
- Enhanced model inference efficiency: Achieved notable reduction in response times.
Deprecation warning: models 1b40a9b479b131e7acb9cab797f929e28eb5dffac86ce1d71d83c564
and 16dde449bdffa504b805508d58e400b18deb4546219e9aeb63edeed2
will be deprecated in release 2024.4.
das-Peak 2.17¶
- New speaker-recognition model with hash
151f12952e9622384073358239301feb1869a3e6eec679953e8f42c5
and a 31.4% relative improvement in False Rejection Rate for Telephone channel. - Deprecation of the
credential/wav
endpoint in favor of the existing and more efficientmodel/{hash}/credential/wav
endpoint.
das-Peak 2.16¶
-
Removed some fields from the API response:
/similarity/credential2wav
: fields removed areauthenticity_reference
,net_speech_duration_reference
, andinput_audio_duration_reference
.
/similarity/credential2credential
: fields removed areauthenticity_reference
,net_speech_duration_reference
,input_audio_duration_reference
,authenticity_to_evaluate
,net_speech_duration_to_evaluate
andinput_audio_duration_to_evaluate
.
/identification/credential2credentials
: fields removed areauthenticity_reference
,net_speech_duration_reference
,input_audio_duration_reference
.
das-Peak 2.14¶
- Improved antispoofing algorithm for telephone channel.
das-Peak 2.13¶
-
Enhanced the antispoofing engine of our voice biometric system to improve the detection and prevention of replay attacks specifically for the microphone channel.
- The enhanced engine has demonstrated a significant increase in detecting both low and high-quality spoofs, with a relative improvement of 37.72% for low-quality spoofs and 80.33% for high-quality spoofs.
Note: low/high quality spoof attacks are specified in the documentation
-
Addition of two new fields to the API's response:
input_audio_duration
: This field provides the duration of the audio received as input by the API.net_speech_duration
: This field provides the duration of the audio after applying our Voice Activity Detection (VAD) technology.
Note: check all the changes in the API's documentation.
das-Peak 2.12¶
- The biometrics model with hash
2b045e0bc0ba5426651d3e4638403da43e5f843afaa32c3fc5773626
has been calibrated for the telephone channel. The new working points have been optimized to balance the system's false positive rate (FPR) and true positive rate (TPR) at a threshold of 0.8. In this new calibration, the FPR has increased from 0.1% to 1%. Under these conditions, the system conversion has improved by a 25% relative improvement. - Reduction of VAD (Voice Activity Detection) inference time by 87% on average.
- Added new
credential2credentials
endpoint for similarity and identifications. The new endpoints are (more details at: https://docs.veridas.com/das-peak/cloud/latest/api/definition/):/similarity/credential2credential
/identification/credential2credentials
- Deprecation announcement: The
credential/wav
endpoint is marked as deprecated, enforcing deprecation for the 2023.7 release. Keep in mind that after the 2023.7 release, it will be necessary to use the endpoint/models/{hash}/credential/wav
.
das-Peak 2.11¶
- Veridas improves its Voice Authentication solution by introducing a new feature that automatically detects audios containing more than one speaker, preventing incorrect registration processes and improving conversion rates. This feature is available in our biometrics model with hash
2b045e0bc0ba5426651d3e4638403da43e5f843afaa32c3fc5773626
.
das-Peak 2.10¶
- Veridas improves its Voice Biometric Authentication solution by enhancing its Voice Activity Detection (VAD) module, the accuracy has been improved by 58.6% compared to the last version.
- Multichannel support (configured to accept a maximum of 2 channels, audio stereo). The effected endpoints are:
- /credential/wav
- /similarity/credential2wav
- /similarity/wav2wav
- /identification/wav2credentials
- /models/{hash}/credential/wav
das-Peak 2.8¶
- This version adds improvements our liveness detection algorithm which is capable of detecting non-genuine voices (prerecorded or synthetic voices). We improved detection of voices coming from high quality speakers by 50.22% and by 36.92% the voices coming from low quality speakers.
- Non-backwards compatible API Change announcement: The /credential/wav endpoint will have hash model ID parameter as mandatory for the 2023 Q1 release.
das-Peak 2.7¶
- New Biometrics model with hash
2b045e0bc0ba5426651d3e4638403da43e5f843afaa32c3fc5773626
and a better performance. - Increased max number of subjects in identification endpoints. The previous N = 100 value has been improved to N = 1000.
das-Peak 2.6¶
- Increased max number of subjects in identification endpoints. The previous N = 10 value has been improved to N = 100.
- The following model with hash '38da15f1b61fb5800c5928f6f1437aed7a0b0e7921fa6bb7852c5783' has been deprecated. The same is not available anymore.
das-Peak 2.4¶
- Updated endpoint '/models/{hash}/credential/wav' to give same AS result than other endpoints.
- Added new voice biometric model with hash '16dde449bdffa504b805508d58e400b18deb4546219e9aeb63edeed2' with last version of antispoofing.
das-Peak 2.3¶
- Different multi-calibration modes: this new functionality allows customers to select between different calibration modes to get the best accuracy of the verification/identification results depending on the use case (telephone audio, lossless audio or non-calibration).
- New endpoint to obtain metadata information of a model given its hash. With this endpoint, it is possible to get some metadata about a specific model.
- New endpoint to obtain the types of calibration that a model supports. With this endpoint, it is possible to know the calibration allowed in a specific model to be used for a specific use case.
- New endpoint to obtain the hash model that was used to create a voice credential.
- Support for a contextual data header: this new header is logged in every query allowing to build statistics and billing segregated by customer, use cases, sub-clients and so on.
das-Peak 2.2¶
This new version of das-Peak incorporates an updated model to detect replay attack spoofs through reproducing voice by smartphone speakers. Briefly, this new version of das-Peak introduces the following changes:
- Pre-recorded voice detection (Replay attack) has significantly improved now reaching 98,62% accuracy. In addition to that this detection method is now represented by a new score that can be optionally activated for backwards compatibility.
das-Peak 2.2.4 (2020-Q1)¶
This new version of das-Peak incorporates new features to work with multiple voice biometric engines and new audio formats. Briefly, this new version of das-Peak introduces the following changes:
- New voice biometric model added which improves accuracy by 26% compared to the previous model. New voice credentials will be generated using the new voice biometric model by default. Previous model can be used as well.
- Capability of processing voice audios with 16Khz sample rate and 16 bits per sample added. das-Peak 2.2.1
This new version of das-Peak incorporates two new endpoints to the API for speaker identification functionalities. Briefly, this new version of das-Peak introduces the following changes:
- Identification endpoint to compare a wav file with a voice credentials list, returning the highest score credential and the rest of scores.
- Identification endpoint to compare a voice credential with a voice credentials list, returning the highest score credential and the rest of scores. das-Peak 1.5
This new version of das-Peak incorporates improved model for speaker verification functionalities. Briefly, this new version of das-Peak introduces the following changes:
- Accuracy of the Voice biometric model improved (EER=4.73 vs EER=3.69, 22% improvement)