New start up warnings if a non-GPU capable plugin is accidentally configured to run on GPU,
Proactively downsample any audio above 16kHz to 16kHz, which is the max sample rate used by any OLIVE plugin. This will reduce memory usage during processing, and also reduce redundant processing during workflows where in the past each plugin in the workflow chain would do its own downsampling to 16 or 8kHz.
Bug Fixes
Fixed a bug with GlobalScorerRequests not honoring passed in audio annotations (regions option) to subselect regions of audio for processing.
Plugins
New Plugins
Enhancement:
enh-mamba-v1.0.0: A very powerful Speech Enhancement plugin that relies on the SE Mamba model architecture.
Speaker Identification, Diarization, Detection (consolidated):
sid-wav2vec-v2.0.0: This new Speaker-centric plugin now combines Speaker Identification (SID), Speaker Diarization (DIA), and Speaker Detection (SDD) into one plugin. This improves maintainability and reduces code duplication on the SRI side, and provides an enhanced user experience by allowing enrollments to be shared for Identification/Verification and Detection tasks, ensuring that both task areas share the latest algorithm and model technology, and allowing the models to share disk and memory space. When the plugins were separate, they each required their own on-disk and in-memory model storage. This single plugin provides highly accurate Speaker Identification on the entire file when tasked as a Global Scorer or Speaker Diarization and Detection when tasked as a Region Scorer. In addition to the combined scoring utility, the plugin is also based on a new and much improved architecture and is significantly more accurate than previous plugins across many data types.
Language Identification, Diarization, Detection (consolidated):
lid-wav2vec-v1.0.0: Similar to the previous bullet, there is now a single Language-centric plugin able to perform the duties that used to be split between Language Identification (LID) and Language Detection (LDD) separate plugins. This is now a single plugin that provides highly accurate Language Identification summarized across an entire file (Global Score) or Language Diarization and Detection timestamped regions within a file (Region Scores). Like the new SID plugin above, recognition performance is also significantly improved with this plugin over previous technologies.
Translation:
tmt-llm-commercial-v1.0.0: Uses an LLM to perform translation of text from one language to another, typically auto-detecting the input language and translating it to English, unlike previous TMT plugins that relied on language-specific domains and put the burden on the user to screen or know the input language, and then task the correct domain. This version of the plugin also provides more human readable output, with capitalization and punctuation. Because it relies on an LLM, it does have fairly heavy resource requirements, and due to the wide coverage, depending on the language and sources, it may not out-perform some of the previous plugins with more specialized training and domains.
Transcription:
asr-end2end-v4.2.0: This plugin was updated to add ASR capabilities (domains) for four new languages: Tagalog, Modern Standard Arabic, Somali, Swahili, and Malay/Indonesian.
General:
Many plugins were updated with bug fixes, minor feature additions, or other revisions. Unlike the plugins above, most of these will not require any change to how they are interacted with, apart from referencing the new plugin names/versions when tasking.
Python API (olivepy)
Improvements
Enhanced client connection timeout behavior
Clients can now pass a timeout to the connect function and if a successful connection cannot be established within the defined timeout in seconds, then a ConnectTimeout is raised, documented here.
Enhanced error reporting when a conditional workflow tasks fails (previously the error message was misplaced in the result as the "jobName")
Additional mechanism for creating workflow helper objects using the client (rather than the workflow definition) using the create_workflow function, documented here.
Support for Audio Converter tasks (such as audio enhancement) added, including the ability to pass the output to downstream plugins. This allows a workflow to, for example, perform Speech Enhancement, and use the enhanced audio output to potentially improve ASR performance.
Improvements
Additional workflow validation logic was added that will proactively confirm conditional plugins and domains are available when the workflow is actualized. Previously, a user would only encounter an error from a missing conditional plugin during workflow analysis.
Raven UI
New Features
Allow enrollment of selected audio regions (rather than entire audio file)
Support for Audio Converter requests (such as audio enhancement)