daniel bigham.ca

Exercise 26: Web queries
October 17, 2008

Summary

To be useful, an in-house AI needs to be able to answer questions that involve getting data from the web.

At its base level, this will be handled as a linguistic transformation. For example:

what temperature is it outside? -> temperature()

A new mapping type called Programs will be defined, which will associate function names with their implementations. For example:

temperature -> Plato.Programs.Temperature.Query

Optionally, a DLL name can be specified:

temperature -> plato.dll: Plato.Programs.Temperature.Query

Arguments can be specified as well:

what's the current temperature in {city} ? -> city_temperature($1)

city_temperature -> Plato.Programs.Temperature.CityQuery

All arguments are of type string and consist of the entity's ID. The function return value is also a string.

Implement the Query and CityQuery functions mentioned here.

Exercise 25: Alternatives to far-field ASR
October 17, 2008

Summary

Although far-field ASR may be an ideal component of an in-house AI, current technology doesn't afford it as an option. This exercise is to determine a viable alternative.

Solution

A wireless microphone is a viable alternative to far-field ASR.

A Bluetooth headset is a good option in terms of cost, non-intrusive design, and range of operation.

IBM research paper on far-field ASR
October 14, 2008

I came across this research paper which was interesting to read through. The highlights were:

1.	The affirmation that far-field speech recognition is challenging.

2.	A mentioned word-error-rate (WER) of around 60%! That would translate to a command-error-rate of < 10%.

3.	Mention of microphone arrays. They appear to help, but not as much as one might hope.

Here's another paper.

older >>