The State of Transcription: Part 1
Should transcription functionality be fully automated, fully human-performed, or a hybrid of these approaches?
In an earlier No Jitter blog post, I discussed the potential benefits that transcription can provide for enterprise communications, such as helping us consume content more quickly (without compromising retention) and being able to better access and analyze the content.
In this and the following two posts, I discuss the state of transcription today. In the final post that I’ll write on this subject, I'll address where I believe the market is going and some key areas of innovation that can help us derive more benefit from recorded audio and video content.
First, let's consider the common use cases for transcription today: voicemail, meeting and medical transcription and closed captioning. The latter two don't directly relate to UC but I'll touch on each briefly because I think there are some interesting lessons to be learned.
Note that this isn't a completely exhaustive list, e.g., I don't address uses of transcription in surveillance applications.
A few years ago, voicemail transcription seemed to garner a fair amount of attention in the media (in UC industry press, anyway). Users touted the benefits of scanning message transcriptions received via SMS or email to get the gist of the original voicemail and determine the priority for required action.
As with speech recognition tasks in the contact center, there is a spectrum of automation applied to this task, enabling vendors to balance the trade-offs between the four key considerations in transcription: cost, accuracy, turnaround time and privacy.
* Fully automated: some providers, including, most notably, Google, use speech recognition and natural language processing algorithms to interpret voicemails. In Google's case, words are color-coded to indicate transcription confidence where lighter text indicates less confidence. While fast and inexpensive, at least on a marginal basis, this approach (somewhat famously) lacks accuracy. According to a recent study by industry analyst Bill Meisel, these fully-automated transcription engines tend to achieve accuracy percentages in the mid-to-high eighties (although these scores might be a bit pessimistic because of the way errors are counted).
* 100% human: another approach relies completely on humans to transcribe messages. Some vendors stream voicemails to transcriptionists as the messages are being left by the caller, allowing for near-real time conversion. While ostensibly solving for turnaround time and accuracy, the downside of this approach is the cost associated with so much human involvement.
* Semi-automated: some vendors (e.g., Nuance through their Jott and SpinVox acquisitions last year) offer a partially automated approach, akin to the agent-assisted IVR processes pioneered over the past few years by start-ups such as Unveil and Spoken. The goal with this approach, of course, is for humans to edit system-generated transcriptions. The human corrections are fed back into the speech recognition engine to improve its accuracy over time. This approach seems to provide the best balance of cost, accuracy and turn-around time for many enterprise use cases.
Although overall voicemail volumes are decreasing in some key enterprise segments, the interest in transcription of these messages is clearly on the rise. Over time, semi-automated transcription is most likely to win out for enterprise communications, where a few dollars a month per user can be easily justified for the productivity and response-time improvements. For consumer applications, the fully automated solutions such as Google Voice voicemail transcription will dominate, as these can scale at the price point that most consumers are comfortable with.
As analyst Dan Miller points out, to drive consumer adoption those marketing these services should attempt to redirect consumer focus from accuracy to usefulness. Consumers will come to accept that good enough is just that, rather than hoping for free or very low cost transcriptions that are 100% accurate.
Since fully-automated voicemail transcriptions won't be very accurate any time soon, marketers must reposition the services. Perhaps it makes more sense to describe these consumer-oriented services under a brand name that speaks to a message's gist or essence rather than a transcription, which implies a certain level of accuracy.
The evolution of the voicemail transcription market teaches us (once again) that managing user perceptions can be as important as the effectiveness of core technology itself.
(to be continued next week)