Text to speech project on microcontroller

[Karamel]

Hi,

I am, currently, taking signals and system couse. So, I want to do this project in next summer with stm32f407 discovery board. (this is exactly hobby project)

Okay.

- I can use fourier transform for generating sound. Because, fourier solution takes too less memory in rom.
- I can find all coefficents of sounds for all alphabet but, is it enough?

For example, "Hello"word says like that "Hel-lo". However, it is not reads like that "H-e-l-l-o". Also alphabeth writes diferent, reads different when there are more letters nearly between eachother. If I must work like syllable by syllable So, there could be too much sound and their datum of coefficent.(it takes in too much space inside rom and I guess, it will not be enough?)

I am confused. Where should I start to research? Now, I have a billion of stupid questions inside my head 

Best Regards,
Karamel 
 

[kevyk]

I worked on the first speech synthesizers back in the 80's and back then, and I think still now, the way quality speech synthesis was accomplished was to store the transitions from phoneme to phoneme (many hundred for English). That will produce a much more natural sound than storing the spectra (or raw signal) of each phoneme itself then trying to string them together. You can experiment by doing it in time domain first.

Search the patent database for speech synthesizers to get some ideas.

[janoc]

Hi,

I am, currently, taking signals and system couse. So, I want to do this project in next summer with stm32f407 discovery board. (this is exactly hobby project)

Okay.

- I can use fourier transform for generating sound. Because, fourier solution takes too less memory in rom.
- I can find all coefficents of sounds for all alphabet but, is it enough?

No, it isn't enough.

BTW, to avoid confusion, you may want to change the title - it is text to speech, not speech to text what you are attempting, if I understand correctly. Speech to text is speech recognition, I doubt you would be able to make that work on a microcontroller beyond a few hardcoded words.

Now to synthesize speech for a given language you should not work with FFT and individual letters, exactly because they are not pronounced the same as they are written. So stringing together sounds of individual letters would not do any good.

Commercial synthesizers use phonemes - all sounds that are used in the given language. Then each word is encoded using these, typically using some rules and/or a large dictionary for all the exceptions.

Look up something called "formant synthesis" - this is the approach that has been used e.g. by the Speak & Spell toy. Good starting point for research could be even this Wikipedia page:
https://en.wikipedia.org/wiki/Speech_synthesis