The uSpeech library provides an interface for voice recognition using the Arduino. It currently produces phonemes, often the library will produce junk phonemes. Please bare with it for the time being. A noise removal function is underway.
The library is quite intensive on the processor. Each sample collection takes about 3.2 milliseconds so pay close attention to the time. The library has been tested on the Arduino Uno (ATMega32). Each signal object uses up 160bytes. No real time scheduler should be used with this.
- Letter based recognition
- Small memory footprint
- Arduino Compatible
- Fixed point arithmetic (not anymore)
- 30% - 40% accuracy if based on phonemes, up to 80% if based on words.
- Plugs directly into an
Head over to the wiki and you will find most of the documentation required.
See installation section of the wiki.
The library utilizes a special algorithm to enable speech detection. First the complexity of the signal is determined by taking the absolute derivative of the signal multiplying it by a fixed point saclar and then dividing it by the absolute integral of the signal. Consonants (other than R,L,N and M) have a value above 40 and vowels have a value below 40. Consonants, they can be divided into frictaves and plosives. Plosives are like p or b whereas frictaves are like s or z. Generally each band of the complexity coeficient (abs derivative over abs integral) can be matched to a small set of frictaves and plosives. The signal determines if it is a plosive or a frictave by watching the length of the utterance (plosives occur over short periods while frictaves over long). Finally the most appropriate character is chosen.
Documentation in other languages are welcome. I will be translating to spanish and chinese (simplified) but the more the merrier. Have a look at the following sections before doing anything:
To get started hacking on this project you should make all changes to the 3.0-workingBranch branch. If you look at the file uspeech.h there are a number of todos listed. Just fork the branch complete a todo and file a pull request.
There is one test: The LED Test