Why isn't speech recognition advancing? [close

Question

What's so difficult about the subject that algorithm designers are having a hard time tackling it?

Is it really that complex?

I'm having a hard time grasping why this topic is so problematic. Can anyone give me an example as to why this is the case?

Answer 1

Auditory processing is a very complex task. Human evolution has produced a system so good that we don't realize how good it is. If three persons are talking to you at the same time you will be able to focus in one signal and discard the others, even if they are louder. Noise is very well discarded too. In fact, if you hear human voice played backwards, the first stages of the auditory system will send this signal to a different processing area than if it is real speech signal, because the system will regard it as "no-voice". This is an example of the outstanding abilities humans have.

Speech recognition advanced quickly from the 70s because researchers were studying the production of voice. This is a simpler system: vocal chords excited or not, resonation of vocal tractus... it is a mechanical system easy to understand. The main product of this approach is the cepstral analysis. This led automatic speech recognition (ASR) to achieve acceptable results. But this is a sub-optimal approach. Noise separation is quite bad, even when it works more or less in clean environments, it is not going to work with loud music in the background, not as humans will.

The optimal approach depends on the understanding of the auditory system. Its first stages in the cochlea, the inferior colliculus... but also the brain is involved. And we don't know so much about this. It is being a difficult change of paradigm.

Professor Hynek Hermansky compared in a paper the current state of the research with when humans wanted to fly. We didn't know what was the secret —The feathers? wings flapping?— until we discovered Bernoulli's force.

Answer 2

The problem is that there are two types of speech recognition engines. Speaker-trained ones such as Dragon are good for dictation. They can recognize almost any spoke text with fairly good accuracy, but require (a) training by the user, and (b) a good microphone.

Speaker-independent speech rec engines are most often used in telephony. They require no "training" by the user, but must know ahead of time exactly what words are expected. The application development effort to create these grammars (and deal with errors) is huge. Telephony is limited to a 4Khz bandwidth due to historical limits in our public phone network. This limited audio quality greatly hampers the speech rec engines' ability to "hear" what people are saying. Digits such as "six" or "seven" contain an ssss sound that is particularly hard for the engines to distinguish. This means that recognizing strings of digits, one of the most basic recognition tasks, is problematic. Add in regional accents, where "nine" is pronounced "nan" in some places, and accuracy really suffers.

The best hope are interfaces that combine graphics and speech rec. Think of an IPhone application that you can control with your voice.

Answer 3

Computers are not even very good at natural language processing to start with. They are great at matching but when it comes to inferring, it gets hairy.

Then, with trying to figure out the same word from hundreds of different accents/inflections and it suddenly doesn't seem so simple.