V‑Blaze 6.1.0
V‑Blaze version 6.1.0 is a major release that provides the following enhancements:
Increased and more consistent performance on systems with multiple NUMA nodes:
Each language model decoder process (and associated memory) is now locked to a single NUMA node.
Language model decode memory usage is re-balanced across NUMA nodes any time a language model is loaded.
JSON changes:
Added the following top-level JSON elements:
started
,ended
: contain timestamps indicating the start and end of processing within the ASR engine. These tags can be useful for measuring latency.streamtags
: contains parameters/tags in the original ASR request. Complex or long parameters (such as scrubconf, subst_rules, sent_rules, etc.) are summarized and confirmed as a count.asr
: contains the version of ASR server being used for the decoding.audiosecs
: contains the total duration (in seconds) of the audio stream.
The
model
element now reflects the actual model used for transcription. This is especially useful if LID is used to automatically select the language model.JSON that contains utterances for multiple channels now orders the utterances based on their start time. Ordering was previously by end time, which resulted in unexpected ordering when one channel overlapped with another.
New options in the vociserver service script:
status -v
now provides full ASR service status informationmemcheck
returns text about general memory health as it pertains to ASR performance (for example, "All nodes healthy")memcheck -v
returns additional performance-related memory details from/proc/buddyinfo
memclean
drops cache and compact memory when necessary for improving ASR performance
Internet licensing improvements:
The worker.cfg can now override settings like numgpus, even if the settings are not defined in the license.
Enabled tracking of number of ASR streams processed for easier verification than relying on call records.
Improved usage tracking based on ASR parameters requested, resulting in a more simplified licensing setup.
Improved robustness of the core StreamListener to avoid edge case hangs and premature connection closures.
Added a default timeout of 30 seconds all native (Python) ASR engine connections.
Increased maximum allowed value for
uttmaxtime
from 120 to 300 seconds.General improvements to system robustness (child process restarts, default timeouts, queues) and transparency (process naming, debug logging).
ASR v6+ Fixed Issue:
Corrected a package dependency that could result in
genderid
being unavailable (voci-server-gid-1.0.0-2)
ASR v5.6.1+ Fixed Issues:
Corrected timestamps and scrubbing for number translation and scrubbing that spans more than one utterance (voci-server-textproc-5.6.1-3).
Fixed rare transcript truncation issues (voci-server-textproc-5.6.1-4).
Improved Spanish language punctuation, capitalization, and number translation (voci-server-textproc-5.6.1-6, voci-server-capitalization-5.6.0-1, voci-langpack-spa-punctuation-2.0.0-2).