searchsound

The SEARCHSOUND project aims to capture and express the disconnected flows of emotion in the internet.

Example output: Processed, Unprocessed

Extenal Libraries

• NLTK
  • Wordnet (for computing word similarities)
• TwitterAPI
• Pianoputer (by Zulko, CC 3.0 Licence, available at https://github.com/Zulko/pianoputer)
  • Pygame
  • Numpy
  • Scipi

How it works

SEARCHSOUND consumes 1% of the recent messages on twitter. It analyses the semantic distance between the words in the tweet and words commonly associated with emotions. It converts this semantic knowledge into a series of tones, and plays them.

It has two modes: short- and long-window

• Short-window
  • Tweets are collected over a window of a few minutes to half-hour. Almost all the input data is used.
• Long-window
  • Tweets are collected over a window of weeks, months, or even a year. Every 8 hour, keywords to track are collected from news services. Tweets with these keywords are tracked, and a mean value for each 2-4 hour period is computed. Chords or a small series of tones is generated for each 2-4 hour period from this mean, and the final 'song' is produced from concatenating these.

But how does it actually work?

It (twitter_stream) uses TwitterAPI and statuses/sample to get a stream consisting of 1% of the Twitter Statuses made every second (specifically, ones made between the 657th and the 666th millisecond). As these tweets come in, it uses NLTK to compute the semantic distance (using LCH similarity) between the words in the tweet and a list words representing an emotion (e.g, happiness) and a list representing its opposite (e.g, sadness). After finding whichever one is closer, it applies a special sigmoid function to the values and saves them in a file.

Then, twitter_player playes those values using the Pianoputer, with each value corresponding to one note.

Current Limitations

Long-window functionality isn't implemented in a usable state, it cannot gather keywords itself nor trim down input data enough to form any long-term analysis of the input data.

Future Plans

• Fully implement long-window functionality
• Consume global searches along with Twitter
• Group all the notes in a short tweet into a chord