Typingpool: Human Audio Transcription Parallelism

theodp writes "Silly rabbit, parallel processing is not just for Big Data! Building on techniques outlined by Andy Baio back in 2008, Wired writer and 20% Doctrine evangelist Ryan Tate has released Ruby-based software called Typingpool to make audio transcriptions easier and cheaper. 'Typingpool chops your audio into small bits and routes them to the labor marketplace Mechanical Turk,' Tate explains to his reporter pals, 'where workers transcribe the bits in parallel. This produces transcripts much faster than any lone transcriber for as little one-eighth what you pay a transcription service. Better still, workers keep 91 percent of the money you spend.' Remember to Use the Force for Good, Tate adds."

Comments

Excuse Me While... I Kiss This Guy.

[cognoscentus]

Seems an interesting idea for long recordings needed in a hurry, but the transcriber will be losing possibly important contextual cues by reducing the length of the utterance. Also, there is a great overhead per-person in terms of manual labour (waiting for audio to buffer, HCI interaction, etc). On the upside, it might be less dull to listen to shorter, more varied recordings than one long one. But it would suck to transcribe half a murder-mystery.

Training

[girlintraining]

The biggest market for audio transcription I'm aware of is in medical transcription, followed by legal. Many of the terms used are not normal english; In fact without a basic understanding of medicine, you could easily confuse one thing for another, with potentially tragic results. This is fine for everyday english, but not in industries where terminology is used that isn't. And that's a lot of it. This would be more useful for something like Siri -- no doubt Apple has humans to process some of the unknowns in the background, and would find a service like this useful, if they don't have something similar already.

Re:Training

[iluvcapra]

I suspect distributing even small, redacted portions of a medical or legal dictation would violate the many confidentiality laws in force in these industries.

I'm a sound editor and from time to time I've toyed with sending certain extremely cretinous jobs to Mechanical Turk, things like cutting silence out of audio recordings (can't always automate this), identifying and synchronizing numerous takes, or going through a scene frame by frame and identifying every frame with a gunshot. It's technically possible but if your project is anything more complicated than the tiniest FunnyOrDie video you're going to be breaching the producer's confidence.

As information technology makes things like Mechanical Turk easier to implement, it makes the information you would send to MT all the more valuable and dangerous to release.

Not sure this is a good idea

[CRCulver]

Five or six years, I transcribed podcasts for Mechanical Turk. Their audio files were already split into shorter passages (3 or 5 minutes, for example). Split them even further, and the transcriber might miss out on the context, which is often vital to knowing what exactly is being jabbered about.

That said, I'm not sure who is the target demographic for this kind of work anymore. Many of the podcasts were on subjects of interest to me, and I was getting about $10/hour from Mechanical Turk, which wasn't bad considering that I was often doing the work from backpacker beach havens around the globe where a couple of hours of work would pay all my expenses for the day. But the last time I had a look at Mechanical Turk, the amount they pay had been heavily reduced. Who wants it now? Even if you are in a cheap third world country, if you have the English skills for such transcriptions, you can surely find better and more dependable work elsewhere.

That's how anime fansubs are done

[Animats]

The volunteer process by which Japanese anime is subtitled within hours of release works a lot like this.