Transcription System: QC & Delivery

This is a series of posts on our human-powered audio transcription system. The following are links to the previous parts: OverviewWorkflowCertificationTranscription & Reviews, Proofreading.

Delivery is the final step of our workflow. Here we perform quality checks on the final transcript that was prepared in the previous step to determine if the transcript is deliverable. When we are satisfied with it we convert the file into MS Word, Adobe PDF, OpenOffiice.org Text file using a template file. Then a notification is sent out to the customer that the transcripts are available for download.

We do two types of quality checks; random sampling and keyword analysis. The keyword analysis helps us spot the least relevant terms in the transcript. The audio around those terms are checked again to ensure that they’re correct. We have developed our own tools which help us do these checks quickly. Typically we don’t have to spend more than the duration of the audio file to complete these checks and deliver the file.

If the quality checks fail — which does happen once in a while — the transcript is proofread once more. We also do a root cause analysis after that to prevent these cases from happening again. That helps us improve our system. Mostly it happens because of poor audio, or difficult accent and/or diction of the speaker.

This post completes this series about our Transcription System. We hope we have given you valuable insight into how we work. We believe we have superior transcription process which performs better and is more scalable than other systems out there. But don’t take our word for it, try out our transcription service today and see for yourself.

Transcription System: Workflow

This is a series on Scribie.com’s audio transcription system. The first part which provides an overview is here

Our workflow consists of five steps.

File Splitting -> Transcription -> Review -> Proofreading -> Delivery

We start by splitting the file into smaller parts. The file is split at the 6 minute boundary which produces one or more files of duration 6 minutes or shorter. This is the first little innovation of our transcription process. File splitting breaks down the work into smaller manageable chunks. It helps in many ways. The file can be worked on parallelly by number of transcribers. A huge amount of effort is not wasted if one part has to be re-done. Additionally, we can track the progress precisely.

Transcription is the typing part. On an average it takes around 15-20 minutes to transcribe a 6 minute file. For a lot of our transcribers–who are mostly home-based freelancers–this is not a huge investment of time. Therefore splitting increases the likely hood that the file will be transcribed quickly. In fact on an average it takes around 1 to 1.5 hours to complete the transcription part of a one hour file!

The accuracy of the transcript is very low at this stage; typically around 50 to 80%. Therefore we do a review. The transcript is checked against the audio and all mistakes are corrected. Time-coding and speaker tracking is also added at this stage. Review usually takes 5 to 8 minutes of effort. But it takes longer for all the parts to get reviewed because we have fewer reviewers than transcribers. This is by design since we promote only our best transcribers to reviewers. The review drastically improves the accuracy.

Once all parts are transcribed and reviewed, we can combine them together and prepare the final transcript. However one more round of review is required here. That’s because, since different parts are worked on by different people, there are bound to be inconsistencies. Proofreading is done by a one person who goes through all the parts together and corrects them. The proofreader is an employee of CGBiz LLC (our company). They are the best of the best we have. We train them and pay them a monthly salary rather than an hourly rate.

The transcript is almost done now. However things might not be perfect even now. The proofreader can make mistakes, some more research may be required for certain terms, etc. So before the delivery we do some random checks. We try to gauge whether the quality is indeed at the level we want it to be. We also use keyword analysis (tf-idf to be precise) to identify out-of-context terms and inconsistencies. We review it again if we are not happy with it. Over time we have found that a small percentage of files require re-review; around 2%. Those are generally the most difficult of files.

Once we are satisfied that the transcript is perfect, as best as it can be, we deliver the file. The file is converted into MS Word, Adobe PDF, OpenOffice Text and plain text formats and we notify the customer that the transcript is available for download.

All of the above happens in 1 day and is managed by our transcription system. We charge only $0.99 per minute of the audio for it. So if you want to get a high quality transcript quickly, please do try out our transcription service today.

The next part of the series talks about the Certification Subsystem.