Announcing Quirkos Transcribe

In qualitative research, most data collection comes from conversations. It's researchers and participants talking, interviewing, running focus groups and sharing verbal information about their lives that makes for rich and revealing discourse.

One of the most requested features from our users has been an automated transcription service. A way that researchers can still be in the cafe, energised from an interview, and turn that precious recording straight into text they can revisit, explore, start to analyse and use for writing up.

It's something I've actually been looking to add to Quirkos for about 6 years now, but couldn't find the right solution. There are plenty of commercial and CAQDAS-connected automated transcription services. But in all that time there were none I felt comfortable using or could recommend when people asked. Usually these were for 3 reasons I hope we've solved: accuracy, data security and cost.

The first is accuracy - it needs to get the words right. And that also means doing a good job at getting strange terminology correct as well, even things like abbreviations (BBC, NPR), names and numbers. Many services we've trailed are OK at this, (Office/Teams/Azure seems particularly bad) but ours is really good. The second part of accuracy which I've seen these systems struggle with is accents - usually these systems are mostly trained to pick up English with an American accent, and any strong accent (even Deep South, Scottish, or Midlands) leads to comical but frustrating errors. Quirkos Transcribe is really good - even with people not speaking in their native tongue, or when their native tongue has a non-American twang. It also supports about a dozen other common languages really well.

Background noise also has a huge effect on accuracy, but most qualitative interviews are recorded in a 'real world' setting - using a phone or dictaphone in a noisy place like a cafe or classroom. Background chatter and music make things difficult for human transcribers, let alone computer ones. But the system we have is by a long way the best with noisy audio.

For a long time, my gold benchmark was a recording I made of a participant with a heavy Scottish accent, which was recorded on a phone (I forgot my dictaphone) in a very noisy cafe. This recording cost me extra to have professionally transcribed as it was so difficult to make out what someone was saying, but Quirkos Transcribe is the first system I have tested that got a good transcription from the recording - as good as the professional service.

The second issue is data security.

We are trying to put the maximum amount of data security in the process with advanced end-to-end and at-rest encryption. This is important, because qualitative data is often confidential, potentially harmful to the participant (and researcher) if leaked, and especially in the audio recording, usually non-anonymised. Most other automated transcription services are not designed like this. The data is sent there-and-back with basic or no encryption. That transcription happens on shared servers with thousands of other users that might try sniffing data. And sometimes the terms and conditions show that the service owners can read, send your data to other companies to process, and even make your data public if they wish. That is not acceptable for most IRBs, ethics boards, or data security reviews when working with confidential data. Our new service keeps your work super-secure, even we can't access your audio or text transcriptions.

The last factor is cost, but to be more specific, I should say value.

There are plenty of commercial automated transcription services, none of them particularly cheap, but still seem like good value compared to human transcription services. However, the automated results still need careful checking - no system, human or computer, is perfectly accurate. So you always need to factor in this time.

And other automated transcription services have quite a high barrier to entry - costing quite a lot for a month or year, and being quite restricted on how much you can transcribe in a period. By hosting our own solution we can offer a low-cost add-on for existing Quirkos Cloud subscribers to directly take hours of audio files, and put transcripts straight in your Quirkos project, or you can export them wherever you like, of course! For a really small monthly fee, you'll be able to transcribe 50 hours a month (not a year like some others!) meaning you can comfortably do all your interviews, personal reflections, verbal diaries, and even draft your papers.

We'll have a much more detailed blog post about the service, how it works and the pricing when it launches. This post also has a lot more information about the caveats to the service - issues with relying on automated transcription, how professional transcribers and self-transcription can be better, and the issues with machine learning and the datasets they are based on. However, we still think this is going to be an amazing tool that will make a huge difference to so many people doing qualitative research.

We aim to have the service live for all Quirkos users in the Spring of 2023, so it's only a few months away! So watch this space, and look here for more detailed information about how to use the automated transcription service.