Today, we will cover simple heuristics to get the most out of your AI. As always, use the AI to do the heavy lifting, and use a human subject matter expert to finesse the results, producing a superior outcome with less time/effort.

The process followed is:

• Transcribe the Gram Marg (literally, ‘road map’ in Hindi) video using Indian English.
• Clean up the transcribed Gram Marg video, and present simple heuristics to the the most out of your AI.
• Translate the video from English to Hindi.

The final result is below. Please note, we have not cleaned up the Hindi result, this post is specifically about the limitations of AI, and where the human subject matter expert needs to take stewardship over the AI.

For context from the website, Gram Marg's main aim is to empower rural India digitally by first and foremost bringing in Internet connectivity thereby empowering rural citizens digitally and in the due course, bringing rural India on board as a major contributor to economic growth and development.

Steps

1. Please direct your browser to videotranslator.ai, and then click on the Login button. Select myTemplate, or your preferred template, and create a new item. Note to follow this visual guide you will require a video component in your template.
2. The original video was sourced from Mozilla’s YouTube channel, and was produced under the auspices of their Equal Rating project. The team behind Gram Marg won the competition, and you should check out their work here - it is a very interesting programme that could improve the lives of millions of people.
3. As a first step, create a new item 'gm_en' and upload the video. This should look like the below when you have finished uploading.
   
4. Now, use Action -> Transcribe, and select Indian English as the dialect of choice.
   
5. This was then corrected manually as described in the heuristics section below. This gives us the same video with English captions, shown below.
6. After that process, the content was translated using the Action -> Translate modal shown below.
   
7. After the translation, we did not fix up the Hindi text. In a normal workflow, you would have a Hindi speaker eyeball and clean up the text for a superior result. The results are shown below.

Heuristics

1. Now, we have the captions - but how good were the actual captions? The following observations can be made:
   1. The quality of the captions was mixed- for example the transcription from 0:20 -> 0:30 was not good.
   2. The AI did not transcribe Gram Marg properly in any instance.
   3. The quality of transcription from 1:24 -> 1:32 was not particularly good. Memorably, the word 'broadcaster' was transcribed to 'Brock Lesnar'.
   4. The narrator uses extra words, in terms of saying 'the' and 'a', and also a few other words which had to be manually removed.
   5. Capitalisation was incorrect in a number of places, and grammar also had to be corrected in some spots.
2. Why did these errors occur, and what can the human subject matter expert do to address these challenges? Some simple fixes, and their context, employed were:
   1. Quite simply, an AI is only as good as its training. The Gram Marg video was explicitly chosen to show the limitations of the AI. The training of the underlying Indian English AI is not very good - so whose fault is that?
   2. The training issue is no error. The simple truth is there is not enough video content in Indian English about this subject matter - the conversion of UHF (Ultra High Frequency) band white spaces into carriers for mobile hotspots. Because the content does not already exist the AI cannot be trained enough to provide a high level of accuracy. Incidentally this is the primary reason why Gram Marg exists, to provide a platform for increasing digital literacy in rural India.
   3. Specific to the broadcaster issue, at a very simple level - how many times is the AI likely to have heard the word 'broadcaster' in Indian English during training - probably not that many. How many times has the AI heard the word 'Brock Lesnar' during training - many more. The AI defaulted to what it knows. More broadly, this is explicitly what we mean by use the human subject matter expert to do the high value tasks.
3. Ok - sometimes the AI gets confused. What about the other errors?
   1. The narrator - Prof Abhay Karandikar - like anyone without extensive media training, speaks in a normal fashion. He is trying to explain the idea, and is thinking about his answers as he speaks. This is normal, but not entirely appropriate for the next step of AI translation, hence needed correction. But why is it not appropriate for AI translation?
   2. People speak in a different way to how they write, especially different to how they write formally. The translation AI is trained primarily on text, and a lot of that text is written in formal styles. Hence, the various idiosyncrasies of how people talk, simply do not translate well into other languages.

Conclusion

The actual fixes are quite simple - do not use the platform as a one and done solution. The platform, and the AI’s are very good at doing the heavy lifting, but in the end, both are simply software.

• Once the AI has finished, have a human subject matter expert eyeball and fix the results. For the majority of content the changes are likely to be minor, because most video’s are made of simple concepts so that they can reach a wide audience.
• Mostly, add grammar and capitalisation. Also remove extra capitalisation.

Video Translator’s platform provides a client with the ability to transcribe and translate video content to/from a number of languages, as discussed here.

Please connect with us on LinkedIn, YouTube or Facebook for any comments, questions, or just to keep up to date with the work we do!

We are very grateful for your support!

The platform is currently in closed beta, where it is being tested. If you are interested in trying out our technology, please drop us an email at hello@videotranslator.ai.