Speech recognition has come a long way and has been one of the leading developments in reach towards the Internet of Things. Now, with Google Home and Amazon’s Alexa, we can control just about any aspect of room ambience without moving an inch. While services like these might lead you to believe that speech recognition has evolved to the point where human transcription just doesn’t make sense anymore, we’re still far from being there yet.
We’re not saying that speech recognition hasn’t grown by leaps and bounds in the past decade or so, far from it, in fact, it’s just that it’s not quite gotten to the point where it can be considered as reliable as the human mind. In other words, if you want accurate transcriptions where context, punctuation and accents are accounted for, going for a transcription writing service is your best bet.
Before we can explore a comparison between these two methods, let us take a look at what both are:
Hiring Professional Transcribers Over Speech Recognition
Manual transcription is fairly straightforward. A person listens to an audio recording and jots down the contents; there isn’t much to explain there.
To understand why human beings are better at this task, we must first understand how we interpret language.
The human mind is a multi-track processor. This means that it can selectively and intuitively utilize multiple processes towards the same end, switching freely between those processes as the need arises. The most powerful tool we have at our disposal when it comes to understanding speech is the flexible use of context.
We can glean details about the subject conversation like several speakers, roles in the conversation, subject matter, and key points, just from listening to the conversation for a small amount of time.
Even if there are noisier sections, we can pause, rewind, and play again until we arrive at a reasonable guess for what is being said.
If that doesn’t work, we can continue further along in the conversation and figure out what might have gone in a particular section through later references.
How Accurate is Speech Recognition
ASR, or Automated Speech Recognition, is the domain of computer science and technological research that deals with getting computers to understand the spoken word. In an ideal scenario, where you’re dealing with clean (noiseless), high fidelity audio, a computer can indeed measure up to human standards and even compete with top-notch transcribers.
Just so you can compare, the human error rate during audio transcription is about four percent, while these artificial intelligence-based technologies usually hover around the five percent mark. That means that they’re almost as good if only slightly worse, than humans at transcribing audio.
The ball game shifts a bit when we talk about LVCSR (Large Vocabulary Continuous Speech Recognition). These applications are categorized by an extremely large vocabulary of words and fluid conversation between multiple speakers. When it comes to these scenarios, the error rate for humans is still at roughly four to five percent, while their technological counterpart measures in at roughly 8 – 10 percent. An accuracy rate of 10 percent means an error every 10 words, which is far beyond the acceptable amount. But real-world applications are much, much worse.
Say you have a phone call or a voice note you’d like transcribed. The background noise can make certain words hard for the speech recognition software to detect, but that’s to be expected. In computer science terms, when you use the word “noise” you’re actually describing different elements that obscure the quality of a particular frame (in the case of audio and video). Noise can come from various sources and increases as you increase volume and distance from the audio capture device. But noise isn’t just an environmental factor, at least not in traditional terms.
You see, when we communicate over telephone lines, the audio signal generated by our voices is actually converted to an electrical signal, which travels across large distances and is then converted back into audio for us to hear. These conversions minimize line loss and maximize distance, allowing us to make phone calls across the world. However, “minimize” is a relative term here; while we can hear and understand a person speaking on a phone call, a computer might not be able to do the same.
There are certain aspects of speech recognition that computers truly excel at. They’re many magnitudes of times faster, and as such, some programs can sift through hundreds of hours of audio in the time, it would take a human being to complete working on a one-hour transcription. However, to mimic a process as complex as understanding human speech, a computer must break it down into individual tasks. Preprocessing, feature extraction, acoustic modelling, language modelling, and decoding are all individual steps in a process designed to achieve the same results as human listening, but that doesn’t mean that both work in the same way.
Why Hiring a Transcription Service is the Better Option
This fluid use of context and a wider understanding of the way objects are related in the world gives human transcribers an edge over conventional computer-based means. Again, even the most advanced neural networks only attempt to mimic our faculty of working with context, and this mimicry is subject to our own understanding of how our mind works, which is admittedly limited.
Systems designed by us to do something our minds do, when we ourselves only have a rudimentary understanding of the internal workings of the human mind, cannot reach the level of accuracy that a professional who has listened to thousands of hours of audio can offer us.
Meanwhile, a computer application will break the process down into multiple simple steps, and by the end of all these processes, arrive at a transcription. While we did mention each of the steps that a hypothetical speech recognition tool could take to recognize patterns in speech, knowing what each of these means isn’t necessary for the purposes of this article.
You need to know that certain aspects of spoken language don’t necessarily translate very well to the machine equivalent of the listening process. The main issues are as follows:
Punctuation
It can be difficult to determine how to punctuate a sentence heard over the phone or out of context, even for human transcriptionists. However, because we have a wider understanding of linguistic rules to work with, we can easily make educated guesses about where to put commas, periods, and semicolons and where to swap out a period for an exclamation mark, for example.
A computer, meanwhile, can only attempt a verbatim transcription, which completely ignores punctuation within a sentence most of the time. What you’re left with is a mess of words separated by no more than a few sparse periods.
Filler Words and Offhand Sounds
These areother aspects of transcription that software fails to reliably account for. Filler words are commonly used unintentionally during the fluid speech, with the most easily explainable reason for their presence being a person struggling to reach for their next word.
A program will not be able to distinguish that a person saying “what’s the word?” is actually trying to recall a word they can’t remember and will instead transcribe that phrase along with what was said previously. It sounds like “oh”s and “ah”s are just as difficult for computers to ignore, and they can make many initial drafts seem like gibberish to the intelligent reader.
Characteristics of speech like a lisp or a stutter are no different from filler words and sounds. In fact, if anything, these are much worse, as they can affect even words the software would normally detect. Tweaking the software to work with these difficulties is often a very complicated process if it’s possible at all, and it might render the software incapable of recognizing normal speech.
Dialect, accents, and other intangible aspects of speech
While it may not have as much bearing on how a particular language is expressed in writing, locality does greatly affect the way we speak any given language. Differences in pronunciation and enunciation between two people within the same locality are often difficult, if not impossible, to account for, so a person speaking in a different dialect may as well be speaking a different language altogether.
Creating software or training an AI agent to understand multiple dialects of the same language is harder than you might think. On some level, the software must determine which particular dialect is being spoken during the feature extraction process. The difficulty of this task is compounded when you have multiple speakers, especially speakers from diverse cultural backgrounds.
The need for edited transcriptions
Edited transcriptions are one possible solution to some of the problems relating to each of the previous aspects of translation. Instead of the verbatim transcription that software generates being the final product, a service provider will often offer some “post-processing” work, in the form of a human being going over the verbatim transcript and fixing minor errors, removing redundancies and filler, and generally improving its readability (by replacing incorrect homophones, for example).
Since this involves essentially hiring a person to proofread a computer-generated translation, it eliminates the “automated” aspect of the process. Instead of using human effort to counter the weaknesses of a computer system (and still arrive at a final product that doesn’t meet the same standards as a professional transcription), why not replace the software application with human effort entirely?
This is not to say that software-generated recordings don’t have their own audience. For some purposes, such as voicemail recordings and general surveillance, automated speech recognition makes perfect sense. In these cases, minor transcription errors can be overlooked in light of the cost-saving that ASR offers. This is also ASR’s greatest saving grace for applications where the volume of incoming data is far too high for a human transcription to be viable.
In cases where high volumes and time-sensitivity factors are considered, edited documents offer the greatest compromise between ASR (which wouldn’t be accurate enough on its own) and professional transcription (which wouldn’t be as cost-effective. For more sensitive applications, especially those in which the contents of a recording are of critical value or where the quality of recordings is inconsistent or cannot be guaranteed, a professionally transcribed document always manages to edge out a computer-generated one.
Hiring Professional Transcribers Vs. Speech Software
As research into natural language processing and speech recognition applications begin to bear greater fruit, we may eventually arrive at a focal point where computer-generated transcriptions are virtually indistinguishable from professionally transcribed ones.
We’re definitely not there yet, though, and given the breadth of problems that computer scientists must tackle to achieve this, it’s safe to say we won’t be there any time soon. While software may achieve great results for extremely specialized applications, a general-use speech recognition software that can reliably get the job done, regardless of the scope of the task, simply doesn’t exist right now.
That’s why we’d suggest that you always opt for professional transcription service over an automatically generated one, especially if you’re invested in maintaining the quality of your final product.
Get in touch and let’s get your transcription project off the blocks right away.
