[smc-discuss] Speech to text tool for Malayalam

[Jerin Philip]

<https://www.aclweb.org/anthology/2020.lrec-1.789>
Hello,

You might want to look into kaldi as Kavya suggested or deepspeech (Mozilla) which is open source with some active community. Training is GPU heavy.

The following publication provides 24 hour single speaker training data (in Malayalam and 2 additional languages), recorded in studio quality.

https://www.aclweb.org/anthology/2020.lrec-1.789

Experiments in paper are for TTS. The training data works both ways. Contact the authors by email, they might have trained models for deepspeech or kaldi. I know they have Hindi ASR (and TTS) models from an earlier publication<https://cvit.iiit.ac.in/research/projects/cvit-projects/facetoface-translation>, unsure if they have Malayalam model but I'm hoping it's in the works. They can possibly point you in the direction of better libraries and resources, as they're actively working in the space. There is active research (and some development) going on as part of a project there.

Disclaimer: I am in some capacity still part of this group.

--
Jerin Philip

________________________________
From: discuss <discuss-bounces at lists.smc.org.in> on behalf of Kavya Manohar <sakhi.kavya at gmail.com>
Sent: Monday, June 22, 2020 7:39:22 AM
To: Discussion list of Swathanthra Malayalam Computing <discuss at lists.smc.org.in>
Subject: Re: [smc-discuss] Speech to text tool for Malayalam

Hi Julin,

We do not have a ready to use, open source  speech to text tool for Malayalam as of now.

But there are works going on in this domain. I am working on a Kaldi<https://kaldi-asr.org/> based solution as an academic researcher. This repo <https://gitlab.com/kavyamanohar/malayalam-spoken-digit-recognizer> demonstrates the development of a spoken digit recognizer for Malayalam. A past attempt on Malayalam ASR using CMUSphinx by Sreenadh is available here<https://github.com/sreecodeslayer/ml-am-lm-cmusphinx>.

SMC has taken initiatives for collecting Malayalam speech corpus<https://msc.smc.org.in> for training general purpose automatic speech recognition (ASR) systems. The audio recorded so far is available here<https://gitlab.com/smc/msc>.
A grapheme to phoneme converter, mlphon<https://gitlab.com/smc/mlphon>  that creates a phonetic lexicon (a dictionary of Malayalam words and its pronunciation) needed for ASR training is another project.  Mlphon is available as a Pypi<https://pypi.org/project/mlphon/> library.

Regards

--
Kavya Manohar
Research Scholar
College of Engineering Trivandrum
https://kavyamanohar.com
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.smc.org.in/pipermail/discuss-smc.org.in/attachments/20200622/05c98b33/attachment-0002.html>