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. It can classify phonemes.
Computer scientists would call it a “feature extraction” toolkit for voice recognition of entire words or sentences. As such it does not have any way of classifying the features of a word. That is left for the user to implement and therefore it can be a bit difficult to use the library.
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 180bytes. 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
The library utilizes a special algorithm to enable speech detection. First the complexity of the signal is determined by the following formula:
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 fricatives and plosives. Plosives are like p or b whereas fricatives are like s or z. Generally each band of the complexity coefficient (abs derivative over abs integral) can be matched to a small set of fricatives and plosives. The signal determines if it is a plosive or a fricative by watching the length of the utterance (plosives occur over short periods while fricatives 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 4.x-workingBranch.