Release date: 19th February 2025

Summary

This page highlights all the changes, new features and bugs addressed within the LumenVox Containers version 5.4.0 release. This change affects Speech products. This version is not available for Voice biometric products, this will be made available in upcoming releases. These release notes also include the following patches:

• asr

  • 5.3.15.3.2

• deployment:

  • 5.3.1

• grammar:

  • 5.3.1

  • 5.3.2

• mrcp-api:

  • 5.3.1

• neural-tts:

  • 5.3.1

This release builds upon the 5.3.0 release - see LumenVox Containers Release Notes 5.3.0.

Highlights

1. LumenVox now supports .gsm audio format
2. Enable TLS communication to Rabbit MQ (Apple request)
3. Support added for X509 for Mongo authentication with TLS encrypted passphrase (Apple request)
4. Mongo sharding has been implemented (Apple request)
5. Prometheus metrics added for MRCP API
6. Enhancements made to improve performance for enhanced transcription
7. Grammar loading speed improved and ripcord implemented to provide error for grammars exceeding pre-defined limit
8. ASR memory growth optimization
9. FinalResults message added when continuous transcription is ended with Finalize API call.
10. Analysis portal now caters for Continuous Transcription interactions
11. New Spanish Encoder model released 4.1.2 to enhance recognition with audio containing single words
12. Add support for en-IE (Ireland) dialect 

 *Note: For TTS, we recommend that text not exceed 4mb (this is roughly 1300 Characters with spaces or around 250 words)

**Note: For Transcription, we recommend that users not exceed 90 minutes of transcribed audio due to gRPC size limits.

What’s new on LumenVox Cloud 5.4.0 

New features

• LumenVox now supports .gsm audio formats

• Enable TLS communication to Rabbit MQ (Apple request)

• Support added for X509 for Mongo authentication with TLS encrypted passphrase (Apple request)

• Mongo sharding has been implemented (Apple request)

• Enhancements made to improve performance for enhanced transcription

• Prometheus metrics added for MRCP API:

  • mrcp_total_requests

  • mrcp_active_requests

  • mrcp_average_request_process_time_dist

  • mrcp_total_responses_returned

  • mrcp_max_calls

  • mrcp_sip_calls

  • mrcp_sip_tcp_connections

  • mrcp_rtsp_calls

  • mrcp_garbage_collection_calls

• Grammar loading speed improved and ripcord implemented to provide error for grammars exceeding pre-defined limit

• ASR memory growth optimization

• FinalResults message added when continuous transcription is ended with Finalize API call.

• Analysis portal now caters for Continuous Transcription interactions

Updates

• New Acoustic model for Spanish created to resolve recognition issues with audio containing single words

• Enabled audio quality tool to run on either GPU or CPU for Apple. New environment variable added- FORCE_CPU_PROCESSING. When set to true, it forces CPU processing even if a GPU is available. In cases where a GPU is unavailable, the speech quality tool will default to CPU processing automatically.

• FinalResults message added when continuous transcription is ended with Finalize API call.

• A memory leak in ASR container resolved.

• New Prometheus metrics added for Apple to track fine-tuned vs regular ASR model usage (asr_fine_tuned_result).

  • FT - if transcription request was processed by the fine-tuned model

  • DNN - if transcription request was processed by the DNN model

  • Error - if there was an error processing interaction

  • Timeout - if there was a decode timeout

• ASR changed to ignore "Garbage" rules as they dynamically create grammars that by default have grammar rules.

• Grammar manager changed to return a grammar load failure if grammars are loaded that are larger than a specified threshold (applied by environment variable if required)

• Grammar service crashes under load using MRCP tests resolved

• Renamed fr-fr Voice Mapping from "Lauren" to "Laurent" 

• Media Server memory leak resolved

• Issue with deployment importing resolved.

• Issue resolved when running ASR/Transcription decodes with audio files longer than 3 minutes in batch mode.

• Continuous transcription issue with barge out errors under load when the Fine tune model is enabled was resolved.

• TTS voice specified in the MRCP header not working resolved.

• Issue when RabbitMQ configuration string was invalid or incorrect causing fatal error resolved.

• Resolved issue with Session management crashes when certain environment variables were empty

• Added support for Spanish post-processing into Spanish ASR/Transcription

• Issue with DTMF input not processed when voice and dtmf grammars are loaded resolved

• Add support for en-IE (Ireland) dialect
  
  

Installation notes

The following helm chart can be used

Helm Chart

Note that for MRCP there is no helm chart but a docker compose file. MRCP will run on its own Docker virtual machine which will integrate into the Kubernetes cluster.

Run the following command helm repo update to update the helm charts

Note: if using TTS we recommend you add the following toggle into the values file legacyEnabled. To enable legacy TTS this must be set to True, and False to enable the new neural TTS. The new neural TTS voices must be loaded in the values file in order for the models to be retrieved from S3

Note if installing from 4.7 or below: There have been helm charts changes - please ensure that if you have custom helm charts that you take note of all the changes before installing/upgrading e.g. licensing has moved from common to global - looking for custom license guid

If installing for MRCP - notes that the conf file settings for MRCP API have been replaced with environment variables e.g. to enable compatibility mode.

Key installation guide changes:

LumenVox now recommends that a minimum of version 1.30 Kubernetes is installed

Upgrade procedures

Upgrade or migration from previous versions is supported. Please contact LumenVox to discuss. See notes above.

Updated API guide

APIs for all speech products available on version 5.4 can be obtained here: https://developer.lumenvox.com/   

Information for voice biometric products relates to version 3.4.0-3.4.3 

Model versions as part of the release

ASR - 4.1.0 (4.1.2 Acoustic model released for Spanish to enhance recognition with audio containing single words)

TTS - 3.0 (Neural TTS) sample rate 24 & 16 - can be down sampled to 8kHz (note change). Legacy TTS models will still run under version 1.0. Further voice enhancements made in version 3.0.1 for energetic call center use).  

VB - 2.1.15

VB incorporates Selene 2.4.3 which was integrated into the Container stack

Model version changes

4.1.2 Acoustic model released for Spanish