Separate words in a sentence.

I am doing speech recognition. I don't have any problem with classifying the words but I don't know how to separate the words in a sentence.

For eg: When I say "How are you?" I want to separate that to "how", "are" and "you" so that I proceed with the processing part (which is pattern recognition).

How would you separate the words in a sentence? I need the code for that.

[Daniel Shub]

What language are you dealing with? What is the approximate SNR of the recordings. Is there substantial reverberation? It is natural speech or where the talkers instructed to speak clearly?

Answer 1

[the cyclist]

The isspace() command might be good enough for what you need. If not, there is also the regexp() command, which you can use to do regular-expression matching.

[Daniel Shub]

I think it is separate words in an audio stream and not a text stream (but I could be wrong).

[the cyclist]

Ah. Upon re-reading, I think you are correct.

[i Venky]

As Daniel is saying, I am talking about audio streaming and not text streaming.

[i Venky]

I am waiting for the answer. Only when I know this I can proceed further.

[Jan]

@i Venky: If you are waiting for an answer, it would be helpful if you reply to Daniels questions in the comment section above.

[aya alofi]

did you solve it ?

[Image Analyst]

aya, did you click the link to the left: Search for tag "speech to text"

You'll probably find something in one of the other answers.

Answer 2

[Walter Roberson]

The solution depends upon the language involved. In "western" languages, word boundaries are marked by short pauses. There are, however, languages which do not use pauses for that purpose, and instead rely upon tone patterns or stress patterns. And watch out for "glottal stops" and clicks, as those may involve pauses that do not mark word boundaries.

Answer 3

[Daniel Shub]

As a first stab you could look at the running RMS power averaged over a short (20 ms) time window. Periods of low power are more likely to be inter-word gaps.

Prms = sqrt(conv(x.^2, ones(Fs*0.01, 1)));

You could also look at the envelope of the speech waveform. You can get the envelope via the Hilbert transform.

Even better would be to use your speech identifier to help identify word boundaries. If the identifier is good, but fails on a chunk of speech, the boundary is probably in the wrong place.