BSS: Speech Analytics

The main articles are:

• Speech technology: On the path from recognition to understanding
• Speech Recognition (Technology, Market)
• Speech synthesis
• Call Center: Purpose, Types, and Tasks

2024: Using generative AI

The updated version of speech analytics from BSS has introduced tools based on generative intelligence. BSS announced this on November 27, 2024.

Access to LLM models for a deeper analysis of dialogs has become available, work with search queries has been simplified and semantic search has been introduced. Innovative plugins for the contact center module (CC) were also implemented, optimizing the work of specialists responsible for checking the quality of calls.

The release introduced an internal voice analytics service that allows you to process dialogs using generative intelligence. Users can access LLM models in single window mode without leaving the system. The service works with both the internal Lama-3 model and external GPT neural networks.

When setting a task, users can ask the LLM model to analyze the entire dialogue, or a certain section of the dialogue, or the speech of a certain speaker. You can also choose from text, numeric, or logical model response formats depending on the task. Next, users need to enter a system prompt that contains a description of the task for LLM. You can test the query you entered before running the task to see which response the LLM will return and make edits if necessary. LLM is able to identify the reasons for the client's negativity in the dialogue, make recommendations to improve the operator's work, calculate the general emotional state of the client and operator, evaluate dialogs and transfer answers to scorecards, identify successful transactions and lost sales in selling scripts.

When working with speech analytics, users use markers - a search query in the system containing certain conditions for filtering (vocabulary, dialogue duration, speech of a certain participant and other parameters). Typically, customers need to create tokens based on lexical dictionaries in one form or another. To create a dictionary from scratch, it is necessary to analyze dialogs in PA and process many text transcriptions to enrich the lexical marker dictionary so that the search query covers as many variations of words or phrases as possible. To simplify interaction with markers, an AI mode was introduced that matches synonyms to phrases specified by the customer. After launching the function, the user independently selects suitable synonyms from the list, and the selected words will immediately turn into marker conditions.

Before running the lexical marker, users can choose the type of search operation: search by simple phrase, by exact phrase or by prefix (word start). Now another type of operation is available - semantic search or search by the vector length of the word. To run a semantic search, you need to specify the desired word and the level of similarity (accuracy) of synonyms. After the start of the operation, those replicas of the dialogue will be highlighted, where, according to the system, there are words that are similar in meaning.

Within the module for the contact center, the tools used by quality specialists who listen to calls, check chats and give ratings have been improved. One of the key improvements is intelligent sampling - an automated procedure for distributing dialogs for quality verification. Previously, the QA manager needed to assign manual evaluation dialogs to supervisors, which required considerable labor, since it was necessary to independently ensure an even number of tasks per employee. Or quality controllers could independently assign dialogues to themselves, which sometimes led to artificial overestimation or underestimation of points due to uneven sampling.

Sampling automatically distributes dialogs to quality controllers. The mechanism will help those contact centers that are focused on fulfilling the requirements for international ISO certification, since it ensures the depth and uniformity of the selection of dialogs, allows you to eliminate the human factor when selecting dialogs for checking quality assessment.

In this release, we presented several important improvements related to the functionality of our RA system. Now voice analytics users can conduct more complex dialog research using LLM models, quickly collect lexical dictionaries for search markers, and schedule the load of quality controllers. We also improved the appearance of our platform: we changed the color scheme, optimized the system menu and added the ability to hide menu blocks. We also increased the number of tools, the color palette of which users can independently customize to their taste. We strive not only to introduce current functions in RA, but also to make interaction with the tool pleasant and intuitive for our clients, - commented Alexander Krushinsky, Director of the Voice Digital Technologies Department of BSS.

2023

Version 2.6 with Evaluation Chart Designer

In version 2.6 of the Voice Analytics system, a scorecard designer and personal accounts for employees of the service quality control group appeared, the functionality of markers was expanded, a plugin was implemented visualization KPI in the system interface. The company BSS announced this on September 19, 2023.

Extending the functionality of markers in BSS speech analytics allows users to hierarchically build queries in token filters. This makes it easier to create new markers and allows you to create a multi-level topic hierarchy to visualize analysis results in reports.

The development of marker functionality allows you to conduct an advanced automatic search for dialogs using various conditions: search for vocabulary when specifying a time sequence, permissible intervals of silence between phrases/words of all participants in dialogs (for example, to search for cases in which operators ask the client to wait on the line and do not return to the client for a long time), as well as use an inaccurate search using which the words specified in the conditions of the marker are found, even if the participant in the dialogue in the process of interaction the word was written with errors.

In the updated version of BSS voice analytics, functionality has appeared for in-depth assessment of the quality of customer service in the contact center. It includes:

• Scorecard Designer. Allows you to create templates for assessing the quality of operators using any assessment criteria and various methods of calculating the final score based on the results of the assessment.
• Assigning tasks to Quality of Service Supervisors. It is a personal account of the chief quality controller (head of the quality control group) and personalized personal accounts of the quality of service controllers.
• Module for reporting and visualization of operator evaluation results using speech analytics scorecards.

Version 2.6 of BSS speech analytics implements the visualization plugin. It allows you to display charts and graphs from different reports on the general dashboard. The number of dashboards is not limited - they can be customized for any indicator or process. As a result, the employee immediately sees the whole "picture" and receives complete information about what is happening in the contact center. Previously, you had to view individual reports, opening them in turn.

5x STT acceleration

BSS accelerated STT in speech analytics by a factor of 5. The company announced this on September 5, 2023.

Acceleration will save customers about 1 million rubles on the purchase of servers or about 520 thousand rubles a year on server rentals at the rate of every 100 operators.

STT (Speech To Text) is a speech recognition subsystem that in the BSS Voice Analytics system is responsible for receiving text transcripts of a telephone conversation or dialogue recorded on a microphone in the service office.

STT is based on a special neural network that translates audio recordings into text. Like any neural network-based technology, STT is very demanding on computing power. For example, in the Voice Analytics system, deployed on 200 operators, about 95% of the server capacity will be serviced by STT.

Accordingly, optimization and acceleration of STT will reduce costs and optimize the implementation and use of voice analytics. - R&D The BSS team actively worked on this issue, relying on its own advanced developments and world research experience.

As a result, we managed to find and implement a "recipe" for the neural network, which made it possible to speed up STT at once 5 times. This means that server capacity with the same amount of load will be required 5 times less.

A 5-fold acceleration is a cautious, "guaranteed" estimate. In some cases, we go to acceleration 13 times. Moreover, our configuration does not require a GPU. This is important, because often customers simply do not have servers with GPUs, and the purchase of such servers can take weeks and months, "commented Alexander Krushinsky, director of the voice digital technologies department at BSS.

He also noted that at the moment he does not know vendors with commensurate sizing.

BSS speech analytics helps you deeply and fully analyze customer interactions at your company's contact center and sales offices. The use of the solution allows you to increase sales and reduce costs, identify effective promotion strategies, increase customer satisfaction, and improve the customer path.

Release of the next version of the system with the addition of the ability to take into account the parameters of scripts

BSS announced on July 10, 2023 the release of the next version of the Speech Analytics system. It implements a number of updated features that will allow users to more easily generate reports on a large number of slices and metrics. And the redesigned search is looking faster and more accurately.

Illustration: YouTube

The BSS Voice Dimension has 12 report templates that are part of the basic delivery. They allow you to display statistics in a tabular and visual view of the quality of operators, the efficiency of handling calls on frequency topics, the percentage of customer satisfaction with service, the state of key performance indicators in the contact center with different degrees of detail over different time intervals.

This version adds the ability to take into account the parameters of scripts, which allows you to create reports with a calculation of the average percentage of development of selected scripts with detail by operators.

In addition, you can install separate plug-ins in BSS speech analytics that provide other advanced features. The plugins are a functional addition to the Voice Analytics system, are included in the basic supply or are installed additionally, at the choice of the customer.

In the release 2.5, updated, as well as completely redesigned previously worked plugins appeared.

In particular, the "Scheduled Reports Plugin" is implemented, which allows you to send the necessary reports to mail on a schedule.

Rewritten Word cloud plugin from scratch. It limits the selection of dialogs on the basis of which a cloud of words (tags) is built, which is useful for projects with large volumes of records. The plugin also allows you to set the logic of searching for words through logical operators and configure a complex search hierarchy at different levels of nesting (dialogue/segment/step/part of speech), select the channel of the dialogue participant in which the words will be searched.

The "Selected Filters" plugin shows all applied filters when all widgets, plugins or dialog tables are displayed. And the Calendar Fast Filters plugin allows you to quickly filter dialogs by time intervals, for example, "yesterday," "last week," and so on.

2021: Version 1.3 with in-depth automatic analysis of all customer service spectra

BSS on October 14, 2021 announced the release of updated versions of the Voice Analytics, Dialog Composer and NLU Suite services.

In Speech Analytics v. 1.3, it became possible to conduct an in-depth automatic analysis of all customer service spectra accessing the contact center or sales office:

1. The emerging tool for denormalizing text transcripts of voice negotiations allows you to see transcribed dialogs in a readable format: taking into account capital letters in proper names, punctuation marks in sentences, the designation of numerals not in words, as it was before, but in numbers (for example, when pronouncing a phone number or contract number).

2. It became possible to arrange weights for script markers, creating an automatic checklist with an integral assessment of the customer service quality against lexical and quantitative-time markers with different weights. Previously, all markers had the same weight, for this reason the quality of service checklist could not be fully implemented.

3. A new tool for visualizing voice analytics results - Sankey charts - allows you to analyze service processes using lexical and quantitative-time markers, illustrating key steps, the intensity of processes in each area, relationships, areas for optimization in the contact center, sales office and the company as a whole.

Dialog Composer allows you to develop dialog applications for interaction through phone, chat, SMS, mobile devices or social networks. Supports open standards (VXML, SIP, MRCP, REST), real-time testing and debugging.

Starting with v. 2.4, Dialog Composer allows you to optimize the architecture of voice and text scripting applications using the Decision maker function.

Now any scenario is easy to configure in such a way that the bot can:

• Define changes to the theme in the middle of the dialog.
• Remember the status of the dialogue on the previous topic.
• Return to the previous topic at the client's request.
• do not ask additional questions if the required information is available (without additional settings).

The NLU Suite tool has the ability to cluster with a given level of similarity, view similar replicas, and then assign a Intenta to a filtered group. This allows you to significantly reduce the time of analytics for manual data markup and optimize the process of building a taxonomy.