3
$\begingroup$

I am doing an ASR(automatic speech recognition) as master thesis on low key dataset. Voice and text data is labelled. There are around 4000 phrases and around 5 hours speech.

I don't have background in speech or signal processing. How huge would be pre processing task? Could someone give me a pointer on how to start with(MOOC..). Is it possible to make something out of this project in 5 months ?

Improve this question
$\endgroup$
4
  • 1
    $\begingroup$ What do you mean by "low key"? When you say voice and text data is labelled, do you mean that you have the text output aligned with the voice already? $\endgroup$
    – Erwan
    Commented Dec 27, 2019 at 15:19
  • $\begingroup$ Low key may not be a right term. What i meant was the size of data is not so much. When I say labelled, it means the for a text is associated with a corresponding voice. $\endgroup$
    – Naveen Gabriel
    Commented Dec 27, 2019 at 15:23
  • $\begingroup$ I'm not expert but as far as I know for training an ASR model you need the audio data and the text aligned as closely as possible, ideally word by word I assume. On the other hand if you had some perfect data it might be too easy, so maybe it's fine this way... Hopefully you'll get an answer from somebody more knowledgeable than me, good luck! $\endgroup$
    – Erwan
    Commented Dec 27, 2019 at 20:45
  • $\begingroup$ @Erwan Wh The text and audio is 100% accurate. How it becomes too easy ? $\endgroup$
    – Naveen Gabriel
    Commented Dec 28, 2019 at 1:56

1 Answer 1

Reset to default
3
+25
$\begingroup$

There would be various ways to approach this problem.

For developing a custom solution the latest research utilizes deep learning such as Recurrent Neural Networks to perform Speech Recognition. Here is one of the most known papers around this approach.

I would suggest you look at DeepSpeech. Either use PyTorch or Tensorflow to start from. Here is a PyTorch example and here is a Tensorflow example.

If you don't know either framework go with PyTorch. It's used more in research and thus will find more examples/guidance in the academic field. I also think Tensorflow has a steeper learning curve partly because they change things a lot (v1.0 vs v2.0)

If you can use an existing solution then I recommend using something like Amazon Transcribe or Google Speech to Text, but from experience, these "off the shelf" services are tailored to overgeneralize and aren't very accurate for specific problems domains. Let me know if I can help further

Improve this answer
$\endgroup$
4
  • $\begingroup$ Thanks a lot. Would love to get in touch with you. May i know your background ? $\endgroup$
    – Naveen Gabriel
    Commented Jan 17, 2020 at 13:03
  • $\begingroup$ @NaveenGabriel you're very welcome. Get in touch with me at isak (at) the link in my profile. Happy to help further $\endgroup$
    – IsakBosman
    Commented Jan 17, 2020 at 16:42
  • $\begingroup$ Hi Isak, I am looking to create a phonetics or pronunciation dictionary for the language. I searched over net and could not find relevant. NeitherI do have knowledge on Somali language as such. I do not know how to deal this . $\endgroup$
    – Naveen Gabriel
    Commented Feb 12, 2020 at 12:59
  • $\begingroup$ @NaveenGabriel I can try and help with that. I suggest posting a new question with all the information and then message me the link. That way it can help others and other experts can also chime in :) $\endgroup$
    – IsakBosman
    Commented Feb 12, 2020 at 17:39

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.