Can machines ever truly replace human transcribers?


These days, reliance on speech recognition technology is high and the error rates even higher, rendering the resulting transcripts virtually useless.  According to research, human transcribers have an error rate of around 4%, while commercially available automated transcription software is found to have an error rate of around 12%.  This difference is largely due to the inability of machines and technology to pick up the nuances of the human language, or to distinguish accents and idioms.  As a result of our Editing Machine Transcription service, here at EliteScribe we have seen more than our fair share of what automated software has (or doesn’t have) to offer and we highlight some of the more common issues below:

Comprehension of accents and dialects

Even if a dialect or accent is unfamiliar, a human audio transcriber can research a questionable word or phrase and come to the right conclusion.  A machine is simply incapable of doing this.  From terminologies to incorrect spelling of foreign words, speech-to-text recognition technology can mix things up that may be beyond understanding, or simply leave gaps where it is unable to find a suitable match. 

Knowledge of industry-specific jargon

Human transcribers are preferred over machines for technical content or medical businesses, or indeed when it comes to legal jargon.  Automated software is not skilled enough to understand similar sounding words which could, certainly in the case of a medical transcription, potentially cost a patient their life.  Take the basic examples of to/two/too and then consider this when relevant to industry-specific terminology.

Filler words, punctuation, background noises and general sound interference

Speech recognition technology has an innate inability to differentiate between a semicolon, comma and other punctuation marks.  Also, speech-to-text technology captures all spoken words, even fillers like uh/ah/erm, along with false starts and incomplete sentences, leading to vague and confusing transcripts.  Human transcription services promise intelligent, edited transcripts.

Quite often, the quality of the audio or video file contains background noises which interfere with the dialogue.  Only a human transcriber is able to rewind, playback, listen carefully ad pick up the pieces to form complete sentences.  Speech recognition technology is not programmed to function in a similar way, resulting in inaccurate transcripts.

Checking for facts and bringing clarity

If there are instances of a speaker saying something out of topic, it is down to the human transcriber to research, check for facts, join all the missing pieces and make sense of the incomplete sentences.  With knowledge and clarity on the topic, a professional transcriber can gather what is being said and produce an accurate transcript.  Automated software cannot fact-check or even attempt to gain clarity of a given topic.

Redundant content and multiple speakers

Time and again, here at EliteScribe, we receive transcriptions for editing that either have no speaker labels, inaccurate labels, or simply dialogue from different speakers running on in the same paragraph.  There is also often the case of redundant content to edit out, whether it is simply an aside (chatter over a tea break) or repetitive phrases.  Only a human can add meaning to the content, keeping the flow going and producing an error-free, well-edited transcript.


In summary, technology may have taken the lead in different spheres of our lives, but human understanding, expertise and knowledge is still a requisite, especially in transcription services.
EliteScribe Ltd