# Using IBM Watson Speech to Text API to translate a ProPublica podcast

An example of using the [Watson Speech to Text API](http://www.ibm.com/smarterplanet/us/en/ibmwatson/developercloud/speech-to-text.html) to translate a podcast from ProPublica: [How a Reporter Pierced the Hype Behind Theranos](http://www.propublica.org/podcast/item/how-a-reporter-pierced-the-hype-behind-theranos/)

This is just a simpler demo of the same technique I demonstrate to make automated video supercuts in this repo: https://github.com/dannguyen/watson-word-watcher

The transcription takes just a few minutes (less if you parallelize the requests to IBM) and is free...but it isn't perfect by any means. It doesn't fare super well on proper nouns:

- Charles **Ornstein's** last name is transcribed as **Orenstein**
- John **Carreyrou**'s last name becomes **John Kerry Roo**
- **Elizabeth Holmes** is **Elisabeth homes**
- **Theranos** isn't popular enough to be a proper noun yet, so it is transcribed, variously, as **their in house**, **their nose**, and **thoroughness**
- in one instance, **their Palo Alto headquarters** is transcribed as **they're palatable headquarters**

On the other hand, **ProPublica**, **Ken Auletta**, **David Boies** are perfectly transcribed. Anyway, [it's decent enough for keyword searching](#file-raw-txt-md).


The [watson-transcribe.py](#file-watson-transcribe-py) script is a Python 3.x script that: 

1. downloads the podcast
2. splits it up into smaller WAV files
3. uploads each WAV file to Watson's API to transcribe
4. saves the JSON response
5. parses the JSON response into two plaintext files


This gist contains what the plaintext files look like:

- [raw.txt.md](#file-raw-txt-md) - just the line-by-line transcription from Watson
- [words.csv](#file-words-csv) - The word-by-word transcription, including the confidence level and the timestamps of the words.