Live Call Translator

📋 What’s about

Live Call Translator provides the possibility of carrying out real-time transcription and translation of what is said and through a text channel allows the agent to write a message in his language. This message will be translated into the customer's language and synthesized vocally.

⚙️ Installation

First, it is necessary to access the machine via ssh and run the following command with the root user to install the package for creating Python virtual environments:
apt-get install python3-venv

The software is executed in the form of an XCALLY plugin, so it’s distributed as a zip archive to be installed directly through the admin interface on XCALLY Motion V3, from the AppZone → Plugins section.

Once the plugin installation file has been loaded, you can proceed with the installation. Then you can configure and create all the elements required by this plugin.

Requirements

In the Actions tab enter the following actions:

These are two custom applications, so drag & drop two 'custom' applications and configure them as follows.

🔌 Plugin configuration

In the general section of the plugin, you will need to enter the URL of the server in use and the 'motion apikey'. In the other sections, you will find the tabs of the individual providers where you will need to enter the required data.

For some providers, it may not be sufficient to request an API key but they need to enable individual transcription, translation and text-to-speech services in their portal

☁️ Configure Cloud Provider

Google

Open your Google Cloud Service accounts, choose a project and create a new service account.
Afterwards, click on the email related to the service account that was created and go to Keys section to create new keys. You can select key type for the private key (among JSON and P12 types). By selecting JSON, the system downloads a JSON file on the computer with information to insert on the configuration. In fact by opening the JSON file you can copy from here (without apices):

• Client email

• Private Key

Enter this information in the respective fields within the plugin interface API Transcription, Translation and Text-To-Speech services must now be enabled. Return to the Google Cloud Console home page and search in the search bar for the three api services and click on enable.

Amazon AWS

To configure this provider you need this information, which you can obtain from your Amazon AWS account:

• Access Key ID

• Secret Access Key

You will also have to enter a region. Please note that the transcription service (streaming in this plugin) is only available in certain regions. Here you can see the regions available for transcription, use one of those that allows streaming transcription.

OpenAI

For OpenAI you only need the API key. From your account access the API section → API keys → Create a new secret key. Copy and paste the API key into the plugin

🔧 Services configurations

In the Transcript tab you can choose the transcript provider and other options:

• Openchannel Account: an Openchannel Account must be indicated

• Show original message: the agent will also be able to see the transcript and not just the translation

• Enable Provider Change For Agent: the agent can change the following transcription parameters: provider, input language and output language

• Skip Agent confirmation on start: Enabling this option will not display the window with the choice of provider, input language(s) and output language.

Other options only available for Google and OpenAI:

• Processed audio length (between 2.0 and 8.0 seconds) ( recommended value 3.3):

  • The recording is divided into smaller audio files, with this option you choose the duration in seconds of these audio files to be sent to the provider for transcription. Reducing the duration reduces the latency between the time you speak and the time the transcription takes place. However, this may reduce the quality of the transcription.

• Consecutive processed audio overlap (between 0.1 and 3.0)( recommended value 0.7):

  • This option affects the overlapping time of two consecutive audio files to be transcribed. Increasing this value will reduce the probability of words being lost between two consecutive audios, but will increase the probability of the final part of the first transcript being repeated in the first part of the next transcript.

Provider features and constraints:

• OpenAI:

  • automatic language recognition.

  • It’s possible to choose a specific language, this could improve the speed and quality of transcription.

  • you will have to indicate an input language for the caller because if the caller has not yet spoken and you want to synthesize a message, you must have a target language.

• Google:

  • It’s possible to choose a maximum of 5 possible input languages to be transcribed. By reducing the number of languages, transcription speed and quality may be improved.

• Amazon AWS:

  • It’s possible to choose only a language.

In the Translate tab you can choose the default translation provider and the language to be translated into (the language who is called):

in the sub-tab of the translation section you can choose the provider to be used for a specific language pair:

Note: it is not necessary to indicate all the pairs as the provider is chosen using this logic if you try to translate a pair not present in the configuration: having set an output language into which to translate the system will select the provider corresponding to the last configuration set which presents the output language considered if it does not exist, the system selects the first language with the same macro-region (if we are looking for en-GB, the system will settle for en-US if it is found). If this search also fails, the system will perform the translation with the default translation provider.

The tts tab is similar to the translation section, there is a default provider. You can also choose the provider, voice type and speech synthesis engine for a specific language:

Finally, there is the LOGS tab where you can view the logs relating to the plugin.

👥 Agent side

the agent cannot change all the configurations that were shown in the previous section. to use the plugin you will have to click on the icon at the bottom on the left.

after answering a call you can click on the call.

if the Skip Agent confirmation on the start option is not enabled, you will see this window:

where the agent will be able to modify, only for the call in progress, the default configurations set by the admin.
If the Skip Agent confirmation on start option is enabled and the Openchannel auto answer is enabled, the openchannel interaction corresponding to the current conversation will be opened and the transcription of the conversation will begin.

The agent will see the start transcript and translation message and then as the customer speaks the transcript will appear on the left in the white background messages and the translation in the agent's language will be in the yellow background messages.

The agent will be able to write messages, also using canned answers, which will be translated into the customer's language and a Text-to-Speech (TTS) speech synthesis will be played in the conversation.

In the case of using the auto-detection mode for choosing the client's language, a feature only available for the OpenAI provider, or using the Google provider with more than one language for transcription, it might be useful for the agent to set the client's language to improve the quality and speed of transcription. To do so, the agent has to type in the asterisk twice and then the language code without specifying the sub-region. For example, if the agent types **EN, lower or upper case is not important, English will be fixed as the customer's language. If he types **IT afterwards he will have fixed Italian as the language for the customer:

For the language codes can be forced by agents, see the tables in the appendix at the end of the document.

Logs

The application stores log files for each component and/or function that is executed in the background. You can find these logs in the Log view, where you can check when a specific log is been updated, download, or delete it.

The logs allow you to check the proper functioning of the application, and to eventually check for errors. As mentioned, the log files are organized by component and by type, and can be of four types:

• Component-combined.log: includes the list of all the complete logs for that component.

• Component-combined.date.log: these files gather the log details for the specified component, including both debug and error messages, and for the day specified in the date part of the name. The logs are rotated daily, meaning that you’ll have a new combined log file for each component for each day. The rotation limit is set to 30 days, meaning that after 30 days the oldest logs will be deleted.

• Component-error.log includes the list of all the error logs for that component.

• Component-error.date.log: these files gather the log details related to the errors only, and for the day specified in the date part of the name. The logs are rotated daily, meaning that you’ll have a new combined log file for each component for each day. The rotation limit is set to 30 days, meaning that after 30 days the oldest logs will be deleted.

• TranslatorService_backen.log: includes the complete logs for the backend service.

• TranslatorService_latency.log: includes information on provider latency times. Average, minimum and maximum latency are provided per call.

TranslatorService_backen.log details:

translateService-error.yyyy-mm-dd.log details:

🔤 Language codes on different providers

Table 1, Google TTS language codes that can be forced by agents

Table 2, OpenAI TTS language codes that can be forced by agents

Table 3, Amazon AWS TTS language codes that can be forced by agents