How not to wreck a nice beach
(Quora)
People nowadays commonly use speech-to-text tools for myriad uses, including writing text messages, entering todo lists, and writing emails. However, increasing numbers of enterprising sorts including just plain folk are starting to warm up to using dictation software for writing longer form content such as journals and even novels.
Dictation software has been generally available for many years, in the form of commercial off the shelf application software, text editors in desktop operating systems, notepads in tablet devices, email clients in mobile phones, and more recently in cloud-based document editors, available to the user from any device that can run a web browser.
But what if you have a large-ish number of audio files? It turns out there isn't a comparably inexpensive and easy way to handle even modestly sized batches of files using those just aforementioned tools. There do exist commercially available alternatives, which will gladly take those batches of audio files off your hands and return high quality text. But these are still rather human labor intensive, which means that they remain expensive.
The typical going rate is about a $1/minute, although with a bit of coding effort (typically meaning slicing files into uniform size work units, farming said work units out to a distributed online on-demand work force such as mechanical turk, then resplicing the files back together) apparently can bring the price down to somewhat, say 70 cents/minute. (The slicing and resplicing is needed for quality assurance as well as to reduce privacy and security issues.) That 70 cents a minute is still rather unaffordable to the typical consumer who is a tad long-winded, much less the odd sort who just happens to have a boat load of such files. Not to mention the concerns related to having unknown on-demand workers hearing an early preview of your memoirs (cf. Long Form Voice Transcription, Wired, April 2016).
Fortunately, an affordable alternative is emerging, that is, if you're willing to write a bit of code. All of the hyperscale cloud api providers have made their deep learning speech-to-text models available for a small fraction of the dollar cost of the aforementioned human-corrected transcription services, and two of them are now appealing enough for use by individuals or nonprofits to apply to long form content. Currently you can get cloud api based speech transcription for prices ranging from 2 to 2.4 cents per minute, with an additional allowance free on a monthly basis. Moreover, they effortlessly handle large batches in a single bound. You upload your files to the cloud, offloading the processing from your laptop. Whereas your desktop software or even embedded operating system software would slow your entire machine down to a crawl during this process you could be processing batches of files in parallel up in the cloud.
Although there is still that small issue of having to write some code in order to get this amazing cost and time reduction. I went and wrote said code.
So this amazing capability is generally available at a reasonable price to anyone with a laptop, internet connection, and credit card with an embarrassingly low limit. But it isn't all peaches and chocolate just yet. There is a still a ways to go before these APIs are reliably useful for converting conversational audio recorded in the wild using mobile phones amidst the usual ambient sounds. Nonetheless, the results are tantalizingly good enough and the pricing dramatically low enough that the time has come to dip my metaphorical toes in the water.
I intend to write more about my experiences using these powerful cloud APIs, so stay tuned if this is of interest to you. Although I myself tend to cringe when the blogger ends their post by promising an enticing followup, which invariably never comes. But I really do sincerely intend to write more on this. Of course, that's what they would all say, I'm sure.
I have already compiled a backlog of interesting results just waiting to be converted into intriguing visuals and wrapped in witty yet insightful prose. Promise.
Actually you don't need to wait until then, as here is some additional background on the project. Here are the results of a preliminary comparison between IBM and Google. Here are results made over reference documents. And here are some results made over audio collected in the wild. There are several conclusions that can be drawn from these findings, which will be the subject of a future post.
In closing, my findings indicate to me that I should be able to extract useful text from my recordings, albeit probably not for the purpose of transcribing meeting minutes. The transcript text generated by the cloud API is largely recognizable to me, although that is often in the same way as my handwriting is recognizable to me because I vaguely remember generating it. The error rates are still too high for these transcripts to allow me to reliably parse meaningful whole sentences much less serve as document of record.
These cloud APIs do not have the same quality as do personalized devices. Voicemail transcripts provided by my phone are subjectively better than what I'm seeing with these cloud APIs. One reason is that because the cloud APIs cannot be trained on my voice, although IBM does allow specifying a custom language model.
The accuracy being provided by the cloud APIs may be suitable for indexing audio files to make them more easily searchable (although phonemes might turn out to be the better data unit for indexing audio than words, e.g., Deepgram). I fully expect the current results to be adequately informative to support meaningful textmining and trend analysis.
Upon reviewing the current batch of results it is fascinating to me how a badly transcribed sequence of just a few words in a row can turn large swaths of text into complete gibberish. If you tell a linguist you intend to recognize speech they might bid you luck, but an evesdropping surfer taking that to mean instead that you intend to wreck a nice beach might wish you harm.
We humans are pretty amazing in our ability to understand each other's muffled speech despite the poor connection and background noise much less the tinnitus and what have you.
Nonetheless, I foresee a not too distant future where it will be possible to cheaply attain human level accuracy transcription of speaker-independent long-form speech from noisy recordings obtained using mobile devices. Culminating in, as Wired's Jesse Jarnow puts it, 'bizarre kinds of unforeseen societal change'.
Until then, court stenographers shouldn't have much to worry about.