This project tests utterances in batch - supporting a variety of data sources and reporting formats

Usage no npm install needed!

<script type="module">
  import bespokenBatchTester from '';


Bespoken Batch Tester

Installation | Execution | GitLab | DataDog | API Docs

This project enables batch testing of utterances for voice experiences.

It leverages Bespoken's Virtual Devices to run large sets of utterances through Alexa, Google Assistant, and other voice platforms.

Getting Started


This package requires Node.js 10 or greater.

To install the Bespoken Batch Tester, just run:

npm install bespoken-batch-tester --save

We recommend creating a new project to store artifacts related to the tester, such as the testing configuration file, CI configuration, and custom source code.

Environment Management

We use dotenv when running locally, which takes environment variables from a local .env file.

To set this up, just make a copy of example.env and name it .env. Replace the values inside there with the correct values for your configuration.

For running with continuous integration (such as Jenkins, Circle CI or Gitlab), these values should instead be set as actual values within the CI environment.

Virtual Device Setup

If you want to use multiple tokens, potentially for different purposes, leverage tags:

  "virtualDevices": {
    "myToken": {
      "tags": ["USAccount"]
    "myOtherToken": {
      "tags": ["UKAccount"],
      "settings": {
        "listener.maxDuration": 5

The tags can then be assigned to a record with record.addDeviceTag:


Only tokens that have that tag (or tags) will be used to process it.

Additionally, per device settings can be set via the settings property. This allows for overriding the behavior of the device as needed.

For more information on the best practices for virtual device management, read our guide here.

Create a Configuration File

Here is a bare minimum configuration file:

  "job": "utterance-tester",
  "sequence": ["open my audio player"],
  "source": "csv-source",
  "sourceFile": "path/to/my/file.csv",
  "virtualDevices": {
    "myVirtualDevice": { "tags": ["my-optional-tags"]}

To get started, cut and paste those settings into a new file, such as batch-test.json.

More information on configuring the batch test is below.

Source Data

Source data by default comes from a CSV file (by default, we look at input/records.csv).

The structure of the file is:

Field Required Description
utterance Yes The utterance to be said to the device
device No The device value corresponds to the tag set on the device. If this value matches a tag, that device will be eligible to process this utterance.
locale No The locale to use for this record. If not set, defaults to 'en-US'
voiceID No The voice ID to use for this record. If not set, defaults to 'en-US-Wavenet-D'
[expected values] No There can zero-to-many expected fields defined - these will automatically be compared to fields on the JSON output

device column

With regard to the device column, if we have two devices, like so:

  "virtualDevices": {
    "device1": { "tags": ["google"]},
    "device2": { "tags": ["alexa"]}

If our record looks like this:

play despacito,alexa,playing the song you requested

The record will only be run with the device2, as it has the tag alexa that corresponds to our DEVICE value.

expected value columns

The expected field values will automatically be compared to the actual value on the response from the device.

For example, if we have a record like this:

play despacito,playing the song you requested

Our response from the virtual device may look like this:

  "transcript": "I don't know that one"

The actual transcript value will be compared to the expected one, and the test will be marked a failure or success based on a partial match comparison (i.e., the actual value must include the expected value though they do not need to be an exact match).

More complex field expressions can be handled with the fields configuration property, described below.

Running the Tester

Once the configuration file is created, just enter:

bbt process batch-test.json

And it will be off and running. In practice, we recommend this not be run locally but in a CI environment.

The tester will create a results.csv file, as well as publish metrics to the configured metrics provider.

In-Depth Configuration

The environment variables store sensitive credentials.

Our configuration file stores information particular to how the tests should run, but of a non-sensitive nature.

An example file:

  "fields": {
    "imageURL": "$.raw.messageBody.directives[1][0].url"
  "interceptor": "./src/my-interceptor",
  "job": "utterance-tester",
  "saveInterval": 300,
  "limit": 5,
  "maxAttempts": 3,
  "backoffTime": 10,
  "metrics": "datadog-metrics",
  "sequence": ["open my audio player"],
  "sequential": false,
  "source": "csv-source",
  "sourceFile": "path/to/my/file.csv",
  "transcript": true,
  "virtualDevices": {
      "tags": ["tag1", "tag2"]
  "virtualDeviceConfig": {
    "phoneNumber": "+1 XXX XXX XXX",
    "twilio_speech_timeout": 2,
    "twilio_timeout": 10,
    "waitTimeInterval": 2000,
    "maxWaitTime": 300000 

Each of the properties is explained below:


Each field represents a value in the JSON response. It will be added to the result output.

If the field also is a column in the CSV file, then the value in the CSV is compared to the value in the actual response.

The fields can be mapped to complex JSON path values, such as: "imageURL": "$.raw.messageBody.directives[1][0].url"

The JSON path will be applied to the actual result from the response. If more than one value matches the JSON path expression, then if ANY of the actual values matches the expected the test will pass.


The interceptor allows for the core behavior of the batch runner to be modified.

There are six main methods:

  • interceptRecord - Called before the record is processed
  • interceptResult - Called before the result is finalized
  • interceptError - Called when the process has an error after the max attempts
  • interceptPreProcess - Called after the records have been loaded but before any of them is processed
  • interceptPostProcess - Called after all the records have been executed
  • interceptRequest - Called before the request is sent to a virtual device

Using interceptRecord, changes can be made to the utterance or the meta data of a record before it is used in a test.

Using interceptResult, changes can be made to the result of processing. This can involve:

  • Adding tags to the result (for use in metrics displays)
  • Changing the success flag based on custom validation logic
  • Adding output fields to the CSV output to provide additional information to report readers

Using interceptError custom code can be called after the max attempts were executed.

Using interceptPreProcess custom code can be called before starting the execution of the records. This can involve:

  • Setting up a local storage
  • Calling an API Be aware that this function will be called when resuming too.

Using interceptPostProcess custom code can be called after the execution of the records. This can involve:

  • Calling an API
  • Saving objects to a local storage.

Using interceptRequest the request can be modified before calling the virtual device. This can involve:

  • Adding more messages (before or after the main one)
  • Adding phrases for better speech recognition
  • Doing operations related to the virtual device about to be used

You can read all about the Interceptor class here:


Time interval where the batch job is saved. It's represented in seconds.

This defaults to 300.


The numbers of records to test during test execution. Very useful when you want to try just a small subset of utterances.


The number of attempts to try if a request has errors. Defaults to 3.


The time in seconds to wait before trying again. Defaults to 10.


We have builtin two classes for metrics: datadog-metrics and cloudwatch-metrics.

This dictates where metrics on the results of the tests are sent.

Additionally, new metric providers can be used by implementing this base class:


For tests in which there are multiple steps required before we do the "official" utterance that is being tested, we can specify them here.

Typically, this would involve launching a skill before saying the specific utterance we want to test, but more complex sequences are possible.


Setting this to true forces records to be processed one after another and not in parallel, regardless of the number of virtual devices that are configured.

This defaults to false.


The source for records. Defaults to csv-source. Additional builtin option is s3-source.

For the csv-source, the source file defaults to input/records.csv. This can be overridden by setting the sourceFile property:


For the s3-source, a sourceBucket must be set. Additionally, AWS credentials must be set in the environment that can access this bucket.

To implement your own custom source, read the API docs.


If set to false, speech-to-text is not performed on the audio response from the device


See the section above for information on configuring virtual devices.


For values other than the default (, set this property.


Allows setting properties to all virtual devices. For example, when using Twilio Virtual Devices, setting the phone_number to call.

Advanced Execution

Resuming A Job

To resume a job that did not complete, due to errors or timeout, simply set the RUN_KEY environment variable.

The run key can be found in the logs for any run - it will appear like this:

BATCH SAVE completed key: 7f6113df3e2af093f095d2d3b2505770d9af1c057b93d0dff378d83c0434ec61

The environment variable can be set locally with:

export RUN_KEY=<RUN_KEY>

It can also be set in Gitlab on the Run Pipeline screen.

Re-printing A Job

CSV reports can be reprinted at any time by running:

bbt reprint <RUN_KEY>

The run key can be found in the logs for any run - it will appear like this:

BATCH SAVE completed key: 7f6113df3e2af093f095d2d3b2505770d9af1c057b93d0dff378d83c0434ec61

Re-processing A Job

Similar to reprinting the CSV results for a job, we can also run a job again applying different tests to the results.

This does NOT call the voice platforms again. Instead, it takes the responses from the platform and pushes them back through the post-processing logic.

This is useful to examine fields that were previously ignored or to change success/failure logic. It allows for fixing errors in the initial analysis without re-doing all the virtual assistant calls.

To use it, enter the following:

bbt reprocess <TEST_FILE> <RUN_KEY>

The TEST_FILE is a the path to the test configuration. The RUN_KEY is the key, shown in the test output, that identifies the job in our storage.

An example call:

bbt reprocess input/bespoken-utterances.json 7f6113df3e2af093f095d2d3b2505770d9af1c057b93d0dff378d83c0434ec61

The run key can be found in the logs for any run - it will appear like this:

BATCH SAVE completed key: 7f6113df3e2af093f095d2d3b2505770d9af1c057b93d0dff378d83c0434ec61

Merge csv results

We take 2 csv files and combine them into one csv file. The result will be located in output/merged.csv by default.


The 2 arguments are optionals, they are by default:

  • ORIGINAL_RESULTS_PATH = ./output/results.csv
  • RERUN_RESULTS_PATH = ./output/rerun.csv

Select output file

You can set the output filename using a flag in any of the commands above

bbt process batch-test.json --output_file your_custom_name

You will find the results in /output/your_custom_name.csv

Output logging

We use overwritten versions of console log, debug, warn and info functions in conjunction with pino.js for logging. This allows us to change the verbosity of our logs (as well as the logs of projects that use the bespoken batch tester) by using the following log levels:

- trace
- debug
- info
- warn
- error
- fatal

On this project, calling console.log is equivalent to calling console.debug, while trace and fatal are not in use. Here's how a typical log looks:

To set a log level, simply set the following env variable: export LOG_LEVEL=debug. If not set, info is assumed.

We can also colorize the outputs by setting the env variable COLORIZE.

Timestamps can be added to the outputs by setting the env variable DISPLAY_LOG_TIME.

Finally, logs can be saved to a file by setting the env variable SAVE_LOG_FILE. The file will be placed in the output folder and named batch-tester.log. It uses pino.js default format.

Type of Utterances

The record class accepts three types of utterances: text, URL and local audio files.

Device Support
Alexa Text, URL, Local Audio File
Google Assistant Text, URL, Local Audio File
Test Robot Text, URL, Local Audio File


  const record = new Record('play a radio station')


  const record = new Record('')

Local audio file

The audio file path should start from the root path of your project

  |- audios/
    |- radio123.wav
|- src/
  |- source.js
  const record = new Record('./subfolder/audio/radio123.wav')

DataDog Configuration

Follow this guide to get started.

Gitlab Configuration

The gitlab configuration is defined by the file .gitlab-ci.yml. The file looks like this:

image: node:10

  - node_modules/

   - npm install
   - npm test
    when: always
      - test_output/report/index.inline.html
      - test-report.xml
      junit: test-report.xml
    - schedules
    - web

When the GitLab Runner is executed, it takes this file and creates a Linux instance with Node.js, executes the commands under the script element, and saves the reports as artifacts.

Setting a schedule

It is very easy to run your end-to-end tests regularly using GitLab. Once your CI file (.gitlab-ci.yml) has been uploaded to the repository just go to "CI/CD => Schedules" from the left menu. Creating a new schedule looks like this:

Test Reporting

We have setup this project to make use of a few different types of reporting to show off what is possible.

The reporting comes in these forms:

  • CSV File that summarizes results of utterance tests
  • Reporting via DataDog
  • Reporting via SQLite and Metabase

Each is discussed in more detail below.

CSV File

The CSV File contains the following output:

Column Description
name The name of the receipt to ask for
transcript The actual response back from Alexa
success Whether or not the test was successful
expectedResponses The possible expected response back from the utterance


DataDog captures metrics related to how all the tests have performed. Each time we run the tests, and when datadog has been set as the metric mechanism to use in the config.json file, we push the result of each test to DataDog.

In general, we are using next metrics:

  • utterance.success
  • utterance.failure

The metrics can be easily reported on through a DataDog Dashboard. They also can be used to setup notifcations when certain conditions are triggered.

Read more about configuring DataDog in our walkthrough.

MySQL and Metabase

MySQL allows for reporting and querying via SQL. It can be combined with Metabase for easy exploration and visualization of the data.

Read more about using these tools in our walkthrough.

Additional Topics