|
|
Pre-Release Notes
Reference Number: AA-02498 Views: 4329 |
0 Rating/ Voters
|
|
Notes on releases not yet GA (General Release)
Important: Message...
Because the development of our DNN ASR engine went hand-in-hand with our move to a container based and cloud native architecture, we have decided to discontinue backporting the new DNN ASR engine to our classic architecture. We released an initial version of our legacy software (version 19.3), that made use of DNN acoustic models for grammar based recognition, but in the future, the DNN ASR engine will be available exclusively in the LumenVox cloud-native architecture. For this reason, we have pulled version 19.3 from general availability. Lumenvox will continue releasing bug and critical fixes for the classic service based architecture, but major feature updates and new languages will be released within the new cloud-native container based architecture. Version 19.3.200 can be obtained by contacting Lumenvox directly, but is not recommended for general production, as no further releases will be made to version 19.3.X related to DNN ASR.
19.3.200 (January 21st, 2022):
Fixes:
- Fixed issue with character encoding generated by the grammar based DNN ASR engine. The results were being returned in the UTF8 character set, but the header specified the results are in ISO-8859-1.
- Fixed inconsistencies that arise when switching back and forth between using the legacy and DNN ASR engines.
- Fixed issue with requests to the legacy ASR engine, for requests using digit-only grammars and not making use of the Legacy digits acoustic model.
- Moved GRAMMAR_ENGINE configuration setting from the SRE section to the GRAMMAR section of the client_property.conf configuration file.
This release currently available for Centos 7 64-bit and Windows.
19.3.100 (January 3rd, 2022):
Improvements and New Features:
- Legacy Grammars: The new end-to-end DNN based recognition engine uses SLMs (Statistical
Language Models) by default. However, in this release the new engine is fully
integrated with the legacy system and can now be used with your legacy grammars.
Grammar-based recognition customers can choose to upgrade the engine to get the
new DNN ASR technology and enjoy the latest in innovation and accuracy, while keeping
the investment in their grammars.
- LumenVox ASR supports both the previous engine and new DNN one at
the same time. You now have the flexibility to choose to use the old or
the new engine for your grammars. You can set
controls that dynamically determine which version is used. If you choose the new engine, you have flexibility to use it with either grammars or SLMs. For more information on these settings see: Using
the DNN ASR Engine with Grammars
- Note
that the new engine ignores phoneme lexicons and in-line phonetic
spelling because the acoustic modeling is no longer phonetic based.
Therefore, we don’t require lexicons in this release when using the new
engine with your grammars. Grammars using lexicons and phonetic spellings
still work, but the alternate pronunciations are ignored. In the future
we plan to return to supporting phoneme lexicons in new and improved ways.
- Windows Support: The DNN ASR service
is only supported on Centos/RHEL 7 64-bit (Linux). However, with
this release:
- The media server, the tuner, and
applications using the API can access the new engine from Windows.
- Additionally, the DNN ASR service
can now be run in a Docker container on any OS that supports running
Linux containers. Many Windows versions support Linux Docker containers. Therefore,
using a Docker installation is a viable option for Windows customers who
want to use the new DNN ASR engine without migrating to Linux.
- Efficiencies & Resources: The DNN ASR engine
takes more resources than the previous engine. Decode operations use,
on average, approximately twice the CPU resources.
- However, the new engine provides
significant improvement in accuracy and lower error rate. The engine
supports operational efficiencies, quantized storage of the language
models, and dynamic usage of ports for call transcription based on load,
thus reducing the bill as compared to using fixed ports.
- Cost Saving Language Licensing: With DNN
recognitions (both grammar based and transcription), licensing is now based
solely on the language, not dialect.
- As an example, a grammar
specifying "en-US" can use an "en-GB" license. The language for a
grammar does not need to specify dialect. Specifying “en”, or “es”, or any other 2 letter language code would suffice.
- To use the DNN ASR engine, you
must install separate language packs. See Language Installation
- Word-Level Confidence Scores: When using
transcription mode, you can obtain more detailed ASR results
in the interpretation section, including word-level confidence scores as
well as word-level timing. See ASR_RESULT_DETAIL_ENABLE
- Wordlists: Use Wordlists for vocabulary customization. The new engine
can accept word and phrase listings, to facilitate recognition of words
that the system has difficulty recognizing, such as pharmaceutical names
or industry jargon – words unique to a domain or business that are not
typically found in a general language model.
- Wordlists are used to augment
language models, not grammars (if you want to augment a grammar you just
edit the grammar). However, to accommodate the creation of wordlists in a
familiar way, we implemented them using a mechanism similar to grammars.
A wordlist definition is straight-forward and looks very similar to a
grammar.
- The words and phrases in the Wordlist
can be given a probability boost.
- See more information to learn how
to create wordlists and how to set a probability boost for wordlists, to enhance language models: Wordlists
This release currently available for Centos 7 64-bit and Windows.
|
|
|