Amazon Elasticsearch Service
Developer Guide (API Version 2015-01-01)

Tutorial: Visualizing Customer Support Calls with Amazon Elasticsearch Service and Kibana

This chapter is a full walkthrough of the following situation: a business receives some number of customer support calls and wants to analyze them. What is the subject of each call? How many were positive? How many were negative? How can managers search or review the the transcripts of these calls?

A manual workflow might involve employees listening to recordings, noting the subject of each call, and deciding whether or not the customer interaction was positive.

Such a process would be extremely labor-intensive. Assuming an average time of 10 minutes per call, each employee could listen to only 48 calls per day. Barring human bias, the data they generate would be highly accurate, but the amount of data would be minimal: just the subject of the call and a Boolean for whether or not the customer was satisfied. Anything more involved, such as a full transcript, would take a huge amount of time.

Using Amazon S3, Amazon Transcribe, Amazon Comprehend, and Amazon Elasticsearch Service (Amazon ES), you can automate a similar process with very little code and end up with much more data. For example, you can get a full transcript of the call, keywords from the transcript, and an overall "sentiment" of the call (positive, negative, neutral, or mixed). Then you can use Elasticsearch and Kibana to search and visualize the data.

While you can use this walkthrough as-is, the intent is to spark ideas about how to enrich your JSON documents before you index them in Amazon ES.

Estimated Costs

In general, performing the steps in this walkthrough should cost less than $2. The walkthrough uses the following resources:

Step 1: Configure Prerequisites

Before proceeding, you must have the following resources.

Prerequisite Description
Amazon S3 Bucket For more information, see Creating a Bucket in the Amazon Simple Storage Service Getting Started Guide.
Amazon ES Domain The destination for data. For more information, see Creating Amazon ES Domains.

If you don't already have these resources, you can create them using the following AWS CLI commands:

aws s3 mb s3://my-transcribe-test --region us-west-2
aws es create-elasticsearch-domain --domain-name my-transcribe-test --elasticsearch-version 6.2 --elasticsearch-cluster-config InstanceType=t2.medium.elasticsearch,InstanceCount=1 --ebs-options EBSEnabled=true,VolumeType=standard,VolumeSize=10 --access-policies '{"Version":"2012-10-17","Statement":[{"Effect":"Allow","Principal":{"AWS":"arn:aws:iam::123456789012:root"},"Action":"es:*","Resource":"arn:aws:es:us-west-2:123456789012:domain/my-transcribe-test/*"}]}' --region us-west-2


These commands use the us-west-2 region, but you can use any region that Amazon Comprehend supports. To learn more, see the AWS General Reference.

Step 2: Copy Sample Code

  1. Copy and paste the following sample code into a new file named

    import boto3 import datetime import json import requests from requests_aws4auth import AWS4Auth import time import urllib2 # Variables to update audio_file_name = '' # For example, 000001.mp3 bucket_name = '' # For example, my-transcribe-test domain = '' # For example, index = 'support-calls' type = 'call' es_region = 'us-west-2' # Upload audio file to S3. s3_client = boto3.client('s3') audio_file = open(audio_file_name, 'r') print('Uploading ' + audio_file_name + '...') response = s3_client.put_object( Body=audio_file, Bucket=bucket_name, Key=audio_file_name, ) response = s3_client.get_bucket_location( Bucket=bucket_name ) bucket_region = response['LocationConstraint'] # Build the URL to the audio file on S3. mp3_uri = 'https://s3-' + bucket_region + '' + bucket_name + '/' + audio_file_name # Start transcription job. transcribe_client = boto3.client('transcribe') print('Starting transcription job...') response = transcribe_client.start_transcription_job( TranscriptionJobName=audio_file_name, LanguageCode='en-US', MediaFormat='mp3', Media={ 'MediaFileUri': mp3_uri }, Settings={ 'ShowSpeakerLabels': True, 'MaxSpeakerLabels': 2 # assumes two people on a phone call } ) # Wait for the transcription job to finish. print('Waiting for job to complete...') while True: response = transcribe_client.get_transcription_job(TranscriptionJobName=audio_file_name) if response['TranscriptionJob']['TranscriptionJobStatus'] in ['COMPLETED', 'FAILED']: break else: print('Still waiting...') time.sleep(10) transcript_uri = response['TranscriptionJob']['Transcript']['TranscriptFileUri'] # Open the JSON file, read it, and get the transcript. response = urllib2.urlopen(transcript_uri) raw_json = loaded_json = json.loads(raw_json) transcript = loaded_json['results']['transcripts'][0]['transcript'] # Send transcript to Comprehend for key phrases and sentiment. comprehend_client = boto3.client('comprehend') # If necessary, trim the transcript. # If the transcript is more than 5 KB, the Comprehend calls fail. if len(transcript) > 5000: trimmed_transcript = transcript[:5000] else: trimmed_transcript = transcript print('Detecting key phrases...') response = comprehend_client.detect_key_phrases( Text=trimmed_transcript, LanguageCode='en' ) keywords = [] for keyword in response['KeyPhrases']: keywords.append(keyword['Text']) print('Detecting sentiment...') response = comprehend_client.detect_sentiment( Text=trimmed_transcript, LanguageCode='en' ) sentiment = response['Sentiment'] # Build the Amazon Elasticsearch Service URL. id = audio_file_name.strip('.mp3') url = domain + '/' + index + '/' + type + '/' + id # Create the JSON document. json_document = {'transcript': transcript, 'keywords': keywords, 'sentiment': sentiment, 'timestamp':} # Provide all details necessary to sign the indexing request. credentials = boto3.Session().get_credentials() awsauth = AWS4Auth(credentials.access_key, credentials.secret_key, es_region, 'es', session_token=credentials.token) # Add explicit header for Elasticsearch 6.x. headers = {'Content-Type': 'application/json'} # Index the document. print('Indexing document...') response = requests.put(url, auth=awsauth, json=json_document, headers=headers) print(response) print(response.json())
  2. Update the initial six variables.

  3. Install the required packages using the following commands:

    pip install boto3 pip install requests pip install requests_aws4auth
  4. Place your MP3 in the same directory as and run the script. A sample output follows:

    $ python Uploading 000001.mp3... Starting transcription job... Waiting for job to complete... Still waiting... Still waiting... Still waiting... Still waiting... Still waiting... Still waiting... Still waiting... Detecting key phrases... Detecting sentiment... Indexing document... <Response [201]> {u'_type': u'call', u'_seq_no': 0, u'_shards': {u'successful': 1, u'failed': 0, u'total': 2}, u'_index': u'support-calls4', u'_version': 1, u'_primary_term': 1, u'result': u'created', u'_id': u'000001'} performs a number of operations:

  1. The script uploads an audio file (in this case, an MP3, but Amazon Transcribe supports several formats) to your S3 bucket.

  2. It sends the audio file's URL to Amazon Transcribe and waits for the transcription job to finish.

    The time to finish the transcription job depends on the length of the audio file. Assume minutes, not seconds.


    To improve the quality of the transcription, you can configure a custom vocabulary for Amazon Transcribe.

  3. After the transcription job finishes, the script extracts the transcript, trims it to 5,000 characters, and sends it to Amazon Comprehend for keyword and sentiment analysis.

  4. Finally, the script adds the full transcript, keywords, sentiment, and current time stamp to a JSON document and indexes it in Amazon ES.


LibriVox has public domain audiobooks that you can use for testing.

(Optional) Step 3: Add Sample Data

If you don't have a bunch of call recordings handy—and who does?—you can index the sample documents in, which are comparable to what produces.

  1. Create a file named

    import boto3 from elasticsearch import Elasticsearch, RequestsHttpConnection import json from requests_aws4auth import AWS4Auth host = '' # For example, region = '' # For example, us-west-2 service = 'es' bulk_file = open('sample-calls.bulk', 'r').read() credentials = boto3.Session().get_credentials() awsauth = AWS4Auth(credentials.access_key, credentials.secret_key, region, service, session_token=credentials.token) es = Elasticsearch( hosts = [{'host': host, 'port': 443}], http_auth = awsauth, use_ssl = True, verify_certs = True, connection_class = RequestsHttpConnection ) response = es.bulk(bulk_file) print(json.dumps(response, indent=2, sort_keys=True))
  2. Update the initial two variables for host and region.

  3. Install the required package using the following command:

    pip install elasticsearch
  4. Download and unzip

  5. Place sample-calls.bulk in the same directory as and run the helper. A sample output follows:

    $ python { "errors": false, "items": [ { "index": { "_id": "1", "_index": "test-data", "_primary_term": 1, "_seq_no": 42, "_shards": { "failed": 0, "successful": 1, "total": 2 }, "_type": "call", "_version": 9, "result": "updated", "status": 200 } }, ... ], "took": 27 }

Step 4: Analyze and Visualize Your Data

Now that you have some data in Amazon ES, you can visualize it using Kibana.

  1. Navigate to

  2. Before you can use Kibana, you need an index pattern. Kibana uses index patterns to narrow your analysis to one or more indices. To match the support-calls index that created, define an index pattern of support*, and then choose Next step.

  3. For Time Filter field name, choose timestamp.

  4. Now you can start creating visualizations. Choose Visualize, and then add a new visualization.

  5. Choose the pie chart and the support* index pattern.

  6. The default visualization is basic, so choose Split Slices to create a more interesting visualization.

    For Aggregation, choose Terms. For Field, choose sentiment.keyword. Then choose Apply changes and Save.

                        Sample configuration for a Kibana pie chart.
  7. Return to the Visualize page, and add another visualization. This time, choose the horizontal bar chart.

  8. Choose Split Series.

    For Aggregation, choose Terms. For Field, choose keywords.keyword and change Size to 20. Then choose Apply Changes and Save.

                        Sample configuration for a Kibana horizontal bar chart.
  9. Return to the Visualize page and add one final visualization, a vertical bar chart.

  10. Choose Split Series. For Aggregation, choose Date Histogram. For Field, choose timestamp and change Interval to Daily.

  11. Choose Metrics & Axes and change Mode to normal.

  12. Choose Apply Changes and Save.

                        Sample configuration for a Kibana vertical bar chart.
  13. Now that you have three visualizations, you can add them to a Kibana dashboard. Choose Dashboard, create a dashboard, and add your visualizations.

                        Sample Kibana dashboard.

Step 5: Clean Up Resources and Next Steps

To avoid unnecessary charges, delete the S3 bucket and Amazon ES domain. To learn more, see Delete a Bucket in the Amazon Simple Storage Service Developer Guide and Delete an Amazon ES Domain in this guide.

Transcripts require much less disk space than MP3 files. You might be able to shorten your MP3 retention window—for example, from three months of call recordings to one month—retain years of transcripts, and still save on storage costs.

You could also automate the transcription process using AWS Step Functions and Lambda, add additional metadata before indexing, or craft more complex visualizations to fit your exact use case.