Typingpool is an app for easily making transcripts of audio using Amazon's labor marketplace, Mechanical Turk.
Typingpool is distributed as a Ruby gem. It is a made up of a handful of scripts for users and a collection of library files for developers.
Typingpool also includes a collection of ERB templates for generating Mechanical Turk assignments and the final transcript HTML file.
Typingpool depends on these command-line tools, which are not included in the gem since they are external to Ruby:
- [ffmpeg] A powerhouse audio/video converter.
- [libmp3lame] An mp3 encoder/decoder, used by ffmpeg.
- [mp3splt] An audio file-splitting utility.
- [mp3wrap] An audio file-merging utility.
After installing the gem and its dependencies, run tp-config from the command line to create your config file (~/.typingpool). At the prompts, you will need to supply your Amazon Web Services Access Key ID and your Amazon Web Services Secret Access key.
The config file is in YAML format and may be customized using any text editor. For more details on configuration options, see the documentation for Typingpool::Config.
A typical workflow will use the bundled scripts in this order:
tp-make -> tp-assign -> [wait] -> tp-review -> tp-finish
tp-review may be called repeatedly, until transcripts for all audio chunks have been processed. Similarly, tp-assign may be called repeatedly, for example to re-assign chunks rejected using tp-review, or to re-assign chunks that have expired.
An alternate workflow would go like this:
tp-make -> [manually upload assignments.csv to Amazon RUI] ->
[wait] -> [approve/reject assignments via RUI] -> tp-collect ->
tp-finish
Typical usage scenario:
tp-make 'Chad Interview' chad1.WMA chad2.WMA --unusual 'Hack Day,
Yahoo' --subtitle 'Phone interview re Yahoo Hack Day'
# => Converting chad1.WMA to mp3
# => Converting chad2.WMA to mp3
# => Merging audio
# => Splitting audio into uniform bits
# => Uploading Chad Interview.00.00.mp3 to
ryantate42.s3.amazonaws.com as Chad
Interview.00.00.33ca7f2cceba9f8031bf4fb7c3f819f4.LHFJEM.mp3
# => Uploading Chad Interview.01.00.mp3 to
ryantate42.s3.amazonaws.com as Chad #
Interview.01.00.33ca7f2cceba9f8031bf4fb7c3f819f4.XMWNYW.mp3
# => Uploading Chad Interview.02.00.mp3 to
ryantate42.s3.amazonaws.com as Chad #
Interview.02.00.33ca7f2cceba9f8031bf4fb7c3f819f4.FNEIWN.mp3
# => ... [snip]
# => Done. Project at:
# => /Users/ryantate/Desktop/Transcripts/Chad Interview
tp-assign 'Chad Interview' interview/nameless --reward 1.00
--deadline 90m --approval 6h --lifetime 2d
# => Figuring out what needs to be assigned
# => 85 assignments total
# => 85 assignments to assign
# => Deleting old assignment HTML from ryantate42.s3.amazonaws.com
# => Uploading assignment HTML to ryantate42.s3.amazonaws.com
# => Assigning
# => Assigned 85 transcription jobs for $85
# => Remaining balance: $115.00
[Wait...]
tp-review 'Chad Interview'
# => Gathering submissions from Amazon
# => Matching submissions with local projects
# =>
# => Transcript for: https://ryantate42.s3.amazonaws.com/
Chad%20Interview.29.00.263d492275a81afb005c8231d8d8afdb.
UEMOCN.mp3
# => Project: Chad Interview: Phone interview re Yahoo Hack Day
# => Submitted at: 2012-08-11 17:00:36 -0700 by A9S0AOAI8HO9P
# =>
# => Chad: ... so it had sort of some geek history. And the
# => weather was really bad. But it was an indoor event,
# => right? So people were staying indoors. And like very
# => early... And there was all this really expensive gear
# => that the BBC had. Like these cameras that guys were like
# => riding around on and stuff, huge sound stage, bigger than
# => the one we had in Sunnyvale.
# =>
# => Two hours into the event, we heard this big lightning
# => strike, because we were up on a hill in London. And all
# => the lights went out and the roof opened up in the
# => building. What we didn't know is the fire supression
# => system in that building which got blown up by the
# => lightning during a fire would cause the roof to open
# => up. So we had all these geeks with equipment and all this
# => BBC equipment and it was literally raining on them.
# =>
# => (A)pprove, (R)eject, (Q)uit, [(S)kip]? (1/20)
a
# => Approved. Chad Interview transcript updated.
# =>
# => Transcript for: https://ryantate42.s3.amazonaws.com/
Chad%20Interview.30.00.263d492275a81afb005c8231d8d8afdb.
RXNKRN.mp3
# => Project: Chad Interview: Phone interview re Yahoo Hack Day
# => Submitted at: 2012-08-11 17:00:58 -0700 by A9S0AOAI8HO9P
# =>
# => Blah blah blah blah okay I am done typing byeeeeeeee
# =>
# => (A)pprove, (R)eject, (Q)uit, [(S)kip]? (2/20)
r
# => Rejection reason, for worker:
There's no transcription at all, just nonsense
# => Rejected
# =>
# => Transcript for...
# => ... [snip]
tp-finish 'Chad Interview'
# => Removing from Amazon
# => Collecting all results
# => Removing HIT 2GKMIKMN9U8PNHKK58NXL3SU4TCBSN (Reviewable)
# => Removing from data/assignment.csv
# => Removing from local cache
# => Removing HIT 2CFX2Q45UUKQ2HXZU8SNV8OG6CQBTC (Assignable)
# => Removing from data/assignment.csv
# => Removing from local cache
# => Removing HIT 294EZZ2MIKMNNDP1LAU8WWWXOEI7O0...
# => ... [snip]
# => Removing Chad Interview.00.00.
263d492275a81afb005c8231d8d8afdb.ORSENE.html from
ryantate42.s3.amazonaws.com
# => Removing Chad Interview.01.00...
# => ... [snip]
# => Removing Chad Interview.00.00.
263d492275a81afb005c8231d8d8afdb.RNTVLN.mp3 from
ryantate42.s3.amazonaws.com
# => Removing Chad Interview.01.00....
# => ... [snip]
The final output of Typingpool is a project directory containing a transcript file.
The transcript file is HTML with audio chunks embedded alongside each associated transcript chunk.
The transcript file is called transcript.html when complete. A partial transcript file is called transcript_in_progress.html.
The project directory also includes supporting files, including a CSV data file used to store raw transcript chunks, Amazon Mechanical Turk HIT information, and other metdata; Javscript code that swaps in Flash players on browsers that don't support mp3 files in audio tags; the original audio files and the audio chunks generated from them; and a CSS file.
The directory is laid out like so:
Chad Interview/
-> transcript.html | transcript_in_progress.html
-> audio/
-> chunks/
-> Chad Interview.00.00.mp3
-> Chad Interview.01.00.mp3
-> ... [snip]
-> originals/
-> chad1.WMA
-> chad2.WMA
-> data/
-> assignment.csv
-> id.txt
-> subtitle.txt
-> etc/
-> audio-compat.js
-> transcript.css
-> About these files - readme.txt
-> player/
-> audio-player.js
-> license.txt
-> player.swf
You may safely edit the files transcript.html, etc/transcript.css, and data/subtitle.txt, and you may safely delete the files in audio/originals and any .txt files in etc/. Editing or deleting other files may interfere with the operation of Typingpool or render the transcript inoperative. Do not edit transcript_in_progress.html as your changes will be overwritten if/when the transcript is next updated.
When you've rejected some submissions in tp-review and need to re-assign these chunks to be transcribed, simply re-run tp-assign with the name of your project. You may select the same template, reward, deadlines, etc., or pick new ones. tp-assign will be careful not to re-assign chunks for which you have approved a transcript, or which are pending on Mechanical Turk.
When some chunks previously assigned via tp-assign have expired without attracting submissions, simply re-run tp-assign as described above to re-assign these chunks. Consider increasing the dollar amount specified in your --reward argument.
When some chunks previously assigned via tp-assign have been submitted by workers but not approved or rejected in time for the approval deadline (assign/approval in your config file or --approval as passed to tp-assign), Mechanical Turk has automatically approved these submissions for you and you'll need to run tp-collect to collect them.
When you want to cancel outstanding assignments, for example because you realize you supplied the wrong parameter to tp-assign, simply run tp-finish with the name of your project. If your assignments have already attracted submissions, you may be prompted to run tp-review first.
When tp-make, tp-assign, or tp-finish unsuccessfully attempts an upload, deletion, or Amazon command, simply re-run the script with the same arguments to re-attempt the upload, deletion or Amazon command. Typingpool carefully records which network operations it is attempting and which network operations have completed. It can robustly handle network errors, including uncaught exceptions.
When you want to preview your assignments, run tp-assign with the --sandbox option and with --qualify 'rejection_rate < 100' (to make sure you qualify to view your own HITs). Then visit http://workersandbox.mturk.com and find your assignments (a seach for "mp3" works if you left mp3 set as a keyword in your config file). When you are done previewing, run tp-finish with the name/path of your project and the --sandbox option.
-
[cache] If the cache file grows too large, you'll need to delete it manually. It may be safely deleted as long as no Typingpool scripts are running. Its location is specified in the 'cache' param in the config file. (The config file is at ~/.typingpool and the cache, by default, is at ~/.typingpool.cache.)
Typingpool takes no steps to limit the size of the cache file. It prunes the cache of project-specific entries when you run tp-finish on a project, but the cache may grow large if you work on many active projects in parallel, or if you fail to run tp-finish on projects when you are done with them. -
[tp-finish] You should run tp-finish PROJECT each time you finish a project, where PROJECT may be either the project name or path. Assuming you have no submissions pending or awaiting approval, this clears all traces of the project from Amazon Mechanical Turk, from Amazon S3 or your SFTP server, and from the local cache. This will keep your local cache from balooning in size and will minimize your S3 charges or SFTP disk usage. It will also help Typingpool scripts run faster by reducing the number of HITs you have on Amazon Mechanical Turk; many Typingpool operations involve iterating through all of your HITs.
-
Run any script with the --help options for further details on how to run the script.
-
See the docs for Typingpool::Config for details of the config file format.
-
See Amazon's Mechanical Turk documentation for guides and overviews on how Mechanical Turk works.
-
See the documentation on ffmpeg and related libraries for clues as to how to make Typingpool support additional file formats. Typingpool can work with any file format that ffmpeg can convert to mp3 (libmp3lame).
Views, used for the final transcript and for rendering HTML assignments for Amazon Mechanical Turk workers, are contained in a series of templates in lib/typingpool/templates, particularly transcript.html.erb and assignment/*. The control layer lives in the App class (lib/typingpool/app.rb) and within the individual scripts. The models constitute the other Typingpool classes, including most importantly and in rough order of importance the Project, Transcript, Amazon, Config and Filer classes (the latter of interest mainly because of Filer::Audio, which handles splitting, merging, and conversion).
The models in particular, along with the App class, are underdeveloped and not particularly clear or fully thought through. The Transcript model, for example, should almost certainly be folded into the Project model. Dividing Project into Project::Local and Project::Remote only makes sense on a superficial level; Project::Remote could probably be its own class or even part of Utility. Amazon will probably be simpler if I can get some patches into RTurk, and Amazon::HIT should probably be integrated more closely with Project.
One of the most frustrating things about the code is that there are so many subtly different ways a "chunk" of a transcript/project is represented: As a simple hash derived from a row in data/assignment.csv within a project folder, as an Amazon::HIT, as a Transcription::Chunk, as an audio file on a remote server, and as a local audio file (which has a different name from the remote file). So in future versions I'll probably reduce the number of different ways to represent a chunk.
Also in the future, it's very likely that App will evolve from a simple collection of class methods into a real class with a simple set of instance methods called in a particular order by a "run" method or similar. Subclasses for particular scripts/commands will then override these methods.
The most comprehensive examples of how the Typingpool classes actually work and interact are the tp-* scripts themselves, in particular tp-make, tp-assign, tp-review, and tp-finish.
More concise examples follow below, to give you a sense of what the various classes actually do:
require 'typingpool'
#new Project instance
project = Typingpool::Project.new('Chad Interview')
#check if project exists on disk
unless project.local
#make a skeleton project folder in Config#transcripts dir
project.create_local
#make subtitle record in project folder
project.local.subtitle = 'Interview about Hack Day Jan 21'
end
id = project.local.id
#Wrap file in Typingpool::Filer
wma = Typingpool::Filer::Audio.new('/foo/bar.wma')
#convert file to mp3
mp3 = wma.to_mp3
other_mp3 = Typingpool::Filer::Audio.new('/foo/bar2.wma').to_mp3
#merge audio
combined_mp3 = Typingpool::Filer::Files::Audio.new([mp3,
other_mp3]).merge(Typingpool::Filer.new('/foo/combined.mp3')
#split audio every 1 minute
chunks = combined_mp3.split('1.00')
#upload mp3s
urls = project.remote.put(chunks.to_streams,
project.create_remote_names(chunks))
#remove mp3s
project.remote.remove_urls(urls)
#new Template instance
template = Typingpool::Template::Assignment.from_config('interview/nameless')
html = template.render({
'audio_url' => urls[0],
'unusual' => ['Hack Day', 'Yahoo', 'Atlassian'],
'chunk_minutes' => 1,
'project_id' => project.local.id
})
question = Typingpool::Amazon::Question.new(urls[0], html)
Typingpool::Amazon.setup
#Assign a transcription job (1 chunk)
hit = Typingpool::Amazon::HIT.create(question, Typingpool::Config.file.assign)
#Find all Typingpool HITs on Amazon Mechanical Turk
all = Typingpool::Amazon::HIT.all
#Find all reviewable Typingpool HITs
reviewable = Typingpool::Amazon::HIT.all_reviewable
#Find all approved Typingpool HITs
approved = Typingpool::Amazon::HIT.all_approved
#Find all HITs for our project
project_hits = Typingpool::Amazon::HIT.all_for_project(project.local.id)
#Filter all HITs (not just Typingpool HITs) arbitrarily
safe_to_delete = Typingpool::Amazon::HIT.all{|hit| hit.ours? && hit.full.expired_and_overdue? }
#Filter all approved HITs arbitrarily
ready_for_judgment = Typingpool::Amazon::HIT.all_reviewable{|hit| hit.submitted? && hit.ours? }
#Approve a HIT
ready_for_judgment[0].at_amazon.approve! #at_amazon is an rturk instance
#Reject a HIT
ready_for_judgment[1].at_amazon.reject!('Your transcription is just random gibberish')
#Delete a HIT from Amazon
safe_to_delete[0].remove_from_amazon
#Get text of transcript chunk (Typingpool::Transcript::Chunk)
transcript_chunk = approved[0].transcript
puts transcript_chunk.body
#Get formmated text of transcript chunk
puts transcript_chunk.body_as_text
#Get transcript chunk as HTML
puts transcript_chunk.body_as_html
#Get transcript chunk metadata
puts "--#{transcript_chunk.url} (audio at #{transcript_chunk.offset})"##Author
Ryan Tate - [email protected]
##License
Copyright (c) 2011-2013 Ryan Tate. Released under the terms of the MIT license. See LICENSE for details.