Twitter Powered Subtitles for Conference Audio/Videos on Youtube

Last week, I wrote a post on hHow people Tweeted Through Carter on “Delivering Digital Britain” at NESTA, where I created a slideshow of tweets posted in response to the NESTA Delivering Britain event/video stream and used them to illustrate the audio recording of Lord Carter’s presentation.

Chatting to @liamgh last week, I mentioned how i was stumped for an easy way to do this. He suggested creating a subtitles feed, and then uploading it to Youtube, along with the audio recording (doh!).

So – here’s a proof of concept result of doing just that:

(Actually, I’m not sure that subtitles/closed captions work in the embedded movies, so you may have to click through: proof of concept use of Tweets as subtitles: Carter at ‘Delivering Digital Britain’, part 1. NB it is possible force embedded videos to show captions, as
Learn More: Showing captions by default
describes. Simply add &cc_load_policy=1 to the embed code; so e.g. in WordPress, you would use something like [youtube= ].

And here’s how I did it…

The first thing to do was check out how Youtube handled subtitles: Getting Started: Adding / Editing captions.

The trick is to upload a textfile with lines that look something like this:
Text shown at 3 min 14.159 sec for an undefined length of time.

Text shown at 2 min 20.25 sec, until 2 min 23.8 sec

Secondly – getting the list of tweets hashtagged with #carter over the period Lord Carter was speaking (i.e. the period covered by the video). For the original proof of concept, I used the tweets from the spreadsheet of scraped tweets that @benosteen grabbed for me, though it later occurred to me I could get the tweets direct from a Twitter search feed (as I’ll show in a minute).

Scraped tweet timestamp

The question now was how to get the timecode required for the subtitles file from the timestamp associated with each tweet. Note here that the timecode is the elapsed time from the start of the video. The solution I came up with was to convert the timestamps to universal time (i.e. seconds since midnight on January 1st 1970) and then find the universal time equivalent of the first tweet subtitle; subtracting this time from the universal time of all the other tweets would give the number of seconds elapsed from the first tweet, which I could convert to the timecode format.

As to how I got the universal time values – I used a Yahoo pipe (Twitter CSV to timeingsd):

At this point, it’s probably worth pointing out that I didn’t actually need to call on @benosteen’s tweetscraper – I could just use the Twitter search API (i.e. the Twitter advanced search feed output) to grab the tweets. How so? Like this:

Looking at the results of this query, we see the timing is a little off – we actually need results from 8.30am, the actual time of the event:

Which is where this comes into play – searching for “older” results:

If you click on “Older” you’ll notice a new argument is introduced into the search results page URL – &page=:

…which means that by selecting appropriate values for rpp= and page= we can tunnel in on the results covering from a particular time by looking at “older” results pages, and grabbing the URL for the page of results covering the time period we want:

Then we can grab the feed:

which gives us something like this:

which we hack a little to get to the right period:

(For more info on the Twitter search API, see the Twitter Search API Documentation wiki.)

NB while I’m at it, note that there’s a corollary hack here that might
come in useful somewhere, or somewhen, else – getting a Twitter search feed into a Google spreadsheet (so we can, for example,process it as a CSV file published from the spreadsheet):


That is:

Okay – back to the main thread – and a tweak to the pipe to let us ingest the feed, rather than the spreadsheet CSV:

Just by the by, we can add a search front end to the pipe if we want:

and construct the Twitter search API URI accordingly:

(The date formatter converts the search date to the format required by the Twitter search API; it was constructed according to PHP: strftime principles.)

Ok – so let’s recap where we’re at – we’ve now got a pipe that will give us universal timecoded tweets (that’s not so far for such a long post to here, is it?!) If we take the JSON feed from the pipe into an HMTL page, we can write a little handler that will produce the subtitle file from it:

Here’s the code to grab the pipe’s JSON output into an HTML file:

var pipeUrl="";

function ousefulLoadPipe(url){

var d=document;
var s=d.createElement('script');

var pipeJSON=url+"&_render=json&_callback=parseJSON";

Here’s the JSON handler:

function parseJSON(json_data){
var caption; var timestamp=0; var mintime=json_data.value.items[0]['datebuilder'].utime;
for (var i=0; itimestamp) mintime=timestamp;

for (var j=json_data.value.items.length-1; j>=0; j--) {
if (j>0) timeEnd=(1*json_data.value.items[j-1]['datebuilder'].utime)-3; else timeEnd=10+1*json_data.value.items[j]['datebuilder'].utime;
if (timeEnd<timestamp) timeEnd=timestamp+2;
var subtitle=timecode+","+timeEnd+"

Here’s the timecode formatter:

//String formatter from:
String.prototype.pad = function(l, s){
return (l -= this.length) > 0 ? (s = new Array(Math.ceil(l / s.length) + 1).join(s)).substr(0, s.length) + this + s.substr(0, l - s.length) : this; };

function getTimeCode(seconds){
var timecode="00:"+(Math.floor(seconds / 60)).toFixed().pad(2, "0") + ":" + (seconds % 60).toFixed().pad(2, "0")+".0";
return timecode;

(I generated the timecodes in part using a string formatter from

Copy and paste the output into a text file and save it with the .sub suffix, to give a file which can then be uploaded to Youtube.

So that’s the subtitle file – how about getting the audio into Youtube? I’d already grabbed an audio recording of Carter’s presentation using Audacity (wiring the “headphones out” to the “microphone in” on my laptop and playing the recording from the NESTA site), so I just clipped the first 10 minutes (I think Youtube limits videos to 10 mins?) and saved the file as a wav file, then imported it into iMovie (thinking I might want to add some images, e.g. from photos of the event on flickr). This crib – iMovie Settings for Upload to YouTube – gave me the settings I needed to export the audio/video from my old copy of iMovie to a file format I could upload to Youtube (I think more recent versions of iMovie support a “Share to Youtube” option?).

I then uploaded this file, along with the subtitles file:

So there we have it: Twitter subtitle/annotations (pulled from a Twitter search feed) to the first part of Lord Carter’s presentation at Delivering Digital Britain…

PS Also on the Twitter front, O’Reilly have started watching Twitter for links to interesting stories, or into particular debates: Twitscan: The Debate over “Open Core”.

Chatting to @cheslincoln the other night, we got into a discussion about whether or not Twitter could be used to support a meaningful discussion or conversation, given the immediacy/short lived nature of tweets and the limited character count. I argued that by linking out to posts to support claims in tweets, “hyper-discussions” were possible. By mining “attention trends” (a term I got from misreading a tweet of Paul Walk’s that scaffold a conversation, it’s possible to create a summary post of a conversation, or argument, like the O’Reilly one?

See also this post from Paul Walk: Anything you quote from Twitter is always out of context.

Simple Embeddable Twitter Map Mashup

Earlier today, I was pondering the Digital Planet Listeners’ map on and the #digitalplanet Twitter feed we have running alongside it:

and I started to wonder whether there was a simple way of generating an embeddable map showing the location of people tweeting given a Twitter feed.

I couldn’t find one offhand, so here’s a simple pipe that will do the job: Simple Tweetmap pipe.

Here’s how it works: start off by grabbing a feed from Twitter, such as a Twitter search feed.

Using a Twitter feed URL as an input to the pipe, grab the feed and then find the Twitter username of each individual from the user’s Twitter URL. So for example, map onto psychemedia.

We now call on another pipe that calls the Twitter API to get personal details for each user who has a Tweet in the feed.

Here’s how that embedded pipe works: Twitter location pipe (it should really be called “Twitter User Details” pipe).

First, construct a URI that ask the Twitter API for the user details associated a particular Twitter username, (e.g. using the construction then pull the data back as a JSON feed. Finally, just make sure only a single record is returned (there should only be one anyway).

So the embedded pipe passes back an annotation to the original feed with user details. One of the user details is the user’s location – so let’s geocode it:

Sending the output of the location finder to item.y:location allows the pipe to produce a well formed KML and geoRSS output, that can be displayed in a map, such as the Yahoo! Pipes output preview map:

We can grab the KML URL from the More Options output and display the feed in an embeddable Google map in the normal way (simply enter the KML URI in the Google maps search box and hit Search):

If you want to embed the map in your own page, just grab the embed code…

To summarise the hack, here’s a quick review of the whole pipe:

So now if you want to plot where people who have tagged their tweets in a particular way are tweeting from, you can do :-)

Next step is to persuade the open2 team to start archiving appropriately tagged tweets and displaying them on a cluster map view over time :-) We could maybe even link in a timeline using something like TimeMap, the MIT Simile timeline and Google Maps integration library…?


Jane’s list of “100+ (E-)Learning Professionals to follow on Twitter” (which includes yours truly, Martin and Grainne from the OpenU :-) has been doing the rounds today, so in partial response to Tony Karrer asking “is there an equivalent to OPML import for twitter for those of us who don’t want to go through the list and add people one at a time?”, I took an alternative route to achieving a similar effect (tracking those 100+ e-learning professionals’ tweets) and put together a Yahoo pipe to produce an aggregated feed – Jane’s edutwitterers pipe

Scrape the page and create a semblance of a feed of the edutwitterers:

Tidy the feed up a bit and make sure we only include items that link to valid twitter RSS feed URLs (note that the title could do with a little more tidying up…) – the regular expression for the link creates the feed URL for each edutwitterer:

Replace each item in the edutwitterers feed with the tweets from that person:

From the pipe, subscribe to the aggregated edutwitters’ feed.

Note, however, that the aggregated feed is a bit slow – it takes time to pull out tweets for each edutwitterer, and there is the potential for feeds being cached all over the place (by Yahoo pipes, by your browser, or whatever you happen to view the pipes output feed etc. etc.)

A more efficient route might be to produce an OPML feed containing links to each edutwitterer’s RSS feed, and then view this as a stream in a Grazr widget.

Creating the OPML file is left as an exercise for the reader (!) – if you do create one, please post a link as a comment or trackback… ;-) Here are three ways I can think of for creating such a file:

  1. add the feed URL for each edutwitter as a separate feed in an Grazr reading list (How to create a Grazr (OPML) reading list). If you don’t like/trust Grazr, try OPML Manager;
  2. build a screenscraper to scrape the usernames and then create an output OPML file automatically;
  3. view source of Jane’s orginal edutwitterers page, cut out the table that lists the edutwitterers, paste the text into a text editor and work some regular ecpression ‘search and replace’ magic; (if you do this, how about posting your recipe/regular expressions somewhere?!;-)

Enough – time to start reading Presentation Zen