I am looking for audio/video samples of conversations between customer service agents and customers.

The background noise and accents are not a problem here so long as the conversation is in English. The dataset must be in an audio/video format and not in written text transcripts.

I expect the conversation between the customer service agents and customers to be revolving more around frequent questions and some new questions.

For Example :- What is the status of my claim number ? (Frequently asked one)

  • Did you find the dataset you were looking for?
    – Rami Segal
    Apr 29, 2019 at 21:39

1 Answer 1


Audio datasets of call center recordings are hard to find as those are usually privately owned and subject to various privacy laws (which differ from one country to another). Unfortunately, it is unclear what are you trying to do with the data and consequently it is hard to give you accurate suggestions that fits your search.

Therefore, depending on your use case you might be able to find -behaviorally- similar datasets. If you are interested in Speaker diarization/ Recogntion over the telephone then datasets like the CallHome database and the CallFriend database -by TalkBank should be sufficient to replicate call center calls (a two-way conversation, sampling rate = 8kHz, telephone noise included)

In case you need to do Speaker verification over the telephone, a dataset of calls-recordings of one individual should be easier to handle and that can be also done using the same -previously mentioned- datasets by separating each sample into two samples (separate speakers, from stereo to mono).

However if you are interested in Speech Recognition over the telephone, then you will need more than a couple of audio recordings. Note that you will need a lot of transcribed datasets (with a sampling rate = 8kHz) and there aren't many out there. This is not an easy task and is indeed an ongoing research area. Moreover, most existing tools for speech recognition are conceived for 16kHz data but there are a couple of decent 8kHz pre-trained models like the CMU Sphinx model here. Alternatively, you can use 16kHz data to train for 8kHz recognition using some transformation like the ones uggested in the following papers:

From the formulation of your question, I suspect you are trying to detect the nature of the call inquiry. This is -to my limited knowledge- a two-steps procedure. First you need to know what is being said? Secondly, what is it about? Of course, you can try using extracted vocational features to predict the nature of the call but I doubt that you will get decent results. For the call natures, I suggest looking for call logs as those are easier to find.

Good luck!

Disclaimer: Please mind the licenses of these datasets and the TalkBank Ground Rules. The TalkBank data is subject to the CC BY-NC-SA 3.0 license and CMU Sphinx is under the BSD-style license.

Not the answer you're looking for? Browse other questions tagged or ask your own question.