In Bahasa Indonesia, we have also similar problem, e.g. Bisa - "venom", Bisa - "able, can", or Panas - "hot", Panas - "high tension".... We usually adding the sentence's context for such kind of problem, or evaluate the speech intensity (energy) , or expand the phonemes into syllable, word or phrase analysis

I think a combination of unsupervised and supervised learning will solve the problem. under the unsupervised aspect, you can develop a codebook for the pronuncation, then apply object recognition during the supervised learning.

There have been some recognition experiments on Yoruba e.g. A Review of Yorùbá Automatic Speech Recognition, Yusuf et al, 2013 IEEE 3rd International Conference on System Engineering and Technology, 19 - 20 Aug. 2013, Shah Alam, Malaysia; CROSS-LANGUAGE MAPPING FOR SMALL-VOCABULARY ASR IN UNDER-RESOURCED LANGUAGES: INVESTIGATING THE IMPACT OF SOURCE LANGUAGE CHOICE, Vakil and Palmer, SLTU 2014. Typically the language model is used to help differentiate between acoustically similar but distinct language contexts. The IARPA Babel programme has been looking at recognition and keyword spotting of low resource languages which covers a wide range of types of languages and should provide some useful information for you. There have been a number of papers published at Interspeech, ICASSP, SLTU and ASRU. The Kaldi toolkit provides a recipe to build a state-of-the-art ASR system for Babel languages. Depending on your experience and the data set(s) available for training your acoustic and language models this may be a good place to start.

I would check the literature for ASR systems for other tonal languages (like Chinese) and see whether you would expect these approaches to carry over to your case.

Here is an example:

Mandarin speech recognition using convolution neural network with augmented tone features 

http://dx.doi.org/10.1109/ISCSLP.2014.6936674

Hi,

you may be interested by this paper :

https://www.researchgate.net/publication/259118437_Automatic_speech_recognition_for_under-resourced_languages_A_survey

also, the following project is maybe of interest for you:

http://alffa.imag.fr

Article Automatic speech recognition for under-resourced languages: A survey

Apply Word Sense Disambiguation using n-gram method to the recognized words from the signal to select the proper word during the final stage of the ASR. All the words (igba) has a different meaning with respect to its preceding or succeeding word(s) with reference to the context.    

A Hidden Markov Model should be able to distinguish between different contexts.Kate provides some good suggestions. I would add to that expanding the HMM beyond typical word pairs and triplets should help. Key, though, is to have enough data for training. In my experience the more ambiguous phonemes are in a language or dialect, the more training data are required to build a language model that has a decent word error rate. The more contextual examples you have, the better your model will perform.

In fact, I am also working in similar lines for speech recognition in Indian Language ("Kannada"). I believe, developing the suitable database is critical. Identifying and tuning the acoustic models to linguistic features would give better results.

Thanks for the discussions.

I think you sould make a systematic search on papers using prosody in ASR,

You should have word models of your laguage in terms of basic units such as phonemes.

You should also build a language model using perhaps  written documents if available

Once You have a lexicon and a phoneic representation of each word, you should collect a corpus of thousands of sentences, perhaps starting with  a small vocbulary and annotate them.

Or just start with a small vocabulary of isolated words.

You should decide the set of acoustic features (including prosodic features)

You may start buiding your acoustic models with HMMs using for example HTK.

Or you can directly build word models

At the beginning you can just try speaker dependent  isolated words. just to assess the value of your features

Good luck

The place to start would be to assemble a corpus of Yoruba speech. You will need it for training and testing purposes. A place to start might be to record a good variety of speakers reading text (such as news articles). Note that the the field has moved on beyond such material, to spontaneous speech and lower-quality speech (say over telephone). But it's good to have a baseline and something on which to work out the peculiarities of ASR for a given language. It's hard to say, but maybe 10 hours of speech would be good. You will need a pronunciation dictionary, and sufficient text to train a language model. If you use news text, there should be plenty. Irrespective of anything else, this will be a very valuable resource.

ASR for tonal languages, such as Mandarin, often models phone and tone together. This is simpler than trying to process the two separately then combining. 

I would suggest finding an already studied language that shares characteristics with Yoruba (like tonality) and assessing the suitability of  their solutions. 

The Besacier et al paper cited above should provide a high-level overview of the issues involved. The languages recently featured in the BABEL program should be a good source of information, though bear in mind that the goal there is key-word spotting, not ASR per se, and so the solutions focus on that.

There are several toolkits available, each with a helpful community. Kaldi provides current DNN modeling techniques. Sphinx supports real-time decoding.  One or the other should fit your needs or style of work. People will likely help you along answering questions.

I see that the LDC has a page on West African languages, and identifies a researcher working on Yoruba. They would certainly know more than someone like me. See https://www.ldc.upenn.edu/sites/www.ldc.upenn.edu/files/west-african-languages.pdf 

Lastly, remember that what you do will be of interest to to others. Be sure to write it up!

I really appreciate all your comments and suggestions. After some literature review that I did, I saw Sphinx-4 as a possible solution to this research work but I'm not sure of its full capability. Can this actually solve the problem?

it seems a good candidate for low resource speech recognition. Look at :

https://www.researchgate.net/publication/271849557_Speech_Acoustic_Unit_Segmentation_Using_Hierarchical_Dirichlet_Processes

There are some other papers (by others and myself) but I have to send them later 

Conference Paper Speech Acoustic Unit Segmentation Using Hierarchical Dirichl...