TTS for Sanskrit

[Avinash L Varna]

Hello all,

I've been investigating cross-lingual TTS techniques to try to adapt available TTS models from other languages for Sanskrit. As part of this investigation, I have been playing around with the Google TTS API and demo - https://cloud.google.com/text-to-speech#section-2.

One experiment I tried is to directly use other Indian languages that mostly have similar phonemes as a substitute for Sanskrit. E.g. use Kannada TTS to generate Sanskrit speech. The results were not bad - definitely intelligible and usable depending on the speed, etc. If you want to try, convert Sanskrit text to Kannada script (e.g. using https://aksharamukha.appspot.com) and paste it in the text section of the demo linked above. Choose Language/locale ಕನ್ನಡ (ಭಾರತ) and one of the available voices. I found that a speed of 0.8 was a good setting, but YMMV.

Since this got tedious after a while, I put together a script to automate it. This weekend, I converted the script into a python package that could be used from other applications - https://github.com/avinashvarna/sanskrit_tts. The package is in a preliminary state, with some defaults that are a personal preference, but could be modified via function arguments. A simple example notebook on using the library is included as well.

For example, I obtained the attached mp3 with the following text from Sanskrit wikipedia (https://sa.wikipedia.org/wiki/कालिदासः#कवेः_देश):

कालिदासस्य जीवनवृत्तिविषये अनेकाः लोकविश्रुतयः अनेके वादाः च सन्ति । केचित् एनं विक्रमादित्यस्य सभायां कविः इति मन्यन्ते । केचित् गुप्तकालीननरेशाणाम् आश्रयं प्राप्तवानिति कथयन्ति । 
धारानगरे भोजराजस्य सभायां कविरत्नपदभूषितः अभूत् इति कथाकोविदाः कथयन्ति। जनश्रुत्यनुसारं बाल्यकाले सः अतीव मूर्खः आसीत् । विद्याधरया सह तस्य विवाहः अभवत् । मूर्खः पतिः इति 
ज्ञात्वा विद्याधरा तं कालीदेव्याः आलयं नीत्वा यावत् सा भवते विद्यां न उपदिशति तावत् भवता ततः बहिः न आगन्तव्यम् इति आदिशत् । ततः पत्न्याः कथनानुसारेण तथैव आचरितः कालिदासः 
कालीदेव्याः वरप्रसादेन विद्वान् अभवत् इति । इयं कथा कालिदासस्य प्रतिभया कविताचातुर्येण च जाता तथ्या तु न इति विदुषां मतिः । यतः जानकीहरणस्य कर्ता राजाकुमारदासः च सप्तमे शतके 
आस्ताम् इति विद्वांसः निश्चितवन्तः । अतः अयं तस्मिन् काले नासीत् इति वक्तुं शक्यते । इतोऽपि कालिदासस्य कविताचातुर्यं निर्दिश्यमानाः अनेकाः कथाः सन्ति ।

There are some systematic errors (e.g. in the pronunciation of visarga following hrasva-i in कविः,पतिः, etc), but the output is probably usable to convert articles/texts to audio for assistive applications, or for listening to articles while commuting/walking/exercising.

Some next steps I am considering:

1. Use this approach to create an initial database of Sanskrit sentences and audio.
   
2. Use database from (1) to create a TTS system that can then be fine-tuned on actual training examples. E.g. maybe fine-tune it on ramayana shloka audio to create a TTS specifically tuned for shlokas as discussed here
3. Over time, add such TTS systems to the sanskrit_tts package
   
4. Maybe use the database from (1) generated with different pitch/speed/etc. settings to bootstrap an ASR system as discussed in the same thread The ASR system could be further fine-tuned on AIR data or others

Any feedback on these steps, and/or the sanskrit_tts library will be appreciated. If you have other ideas for TTS/ASR, please add those as well. Collaborators are always welcome! :)

Regards,

Avinash

[Ganesh S]

नमस्ते अविनाशमहोदय!

एषः तु उत्तमः उपायः अस्ति।

The audio is very good. And most of the pronunciation is correct. I will try out your code soon and see if we can improve in some way.

Have you tried Malayalam and Telugu? I think most of the samskrita aksharANi are in those languages too. If we get good tts from those models too then we can use them too for creating some diversity.

Some thoughts about next steps:

1. We should decide on a prose text corpus to synthesize. VyavahArasahasri text would be a very nice dataset to collect from humans too apart from using TTS. We can also select balamodini kathAH text. Let’s fix on a text corpus that would be approximately 2-4hours and synthesize it. That would be a first annotated samskritam speech dataset using synthetic speech.

2. Can we use Telugu and Malayalam too to add more diversity to the speaker set of the speech corpus? This needs to be explored.

I will devote some time to explore this and get back here.

DhanyavAdaH,

gaNeshaH

--
You received this message because you are subscribed to the Google Groups "sanskrit-programmers" group.
To unsubscribe from this group and stop receiving emails from it, send an email to sanskrit-program...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/sanskrit-programmers/CAALtx9ZiM7i2mYKG0dho5sq16PNoBCOz2sb4bdtGdk7T6-K6jQ%40mail.gmail.com.

--

Ganesh Sivaraman

http://gsivaraman.weebly.com/

[Dhaval Patel]

Dear Avinash, 

The result is absolutely fine. Over and above visargas, the pronunciation of थ sounded like त. 

But these are edge cases. The output was absolurely intelligible and pleasant to hear.

To view this discussion on the web visit https://groups.google.com/d/msgid/sanskrit-programmers/CACsLTJLpLbd0fhvaooZyHmGCRzwmZcyc9K-sFnWq7O3CZN0Z-Q%40mail.gmail.com.

[Gopinath, Raghavendra]

Namaste Avinash, Ganesh,

The audio sounds great.

I built an Android app for Samskrita VyavahArasAhasrI last year end. This app uses Hear2Read TTS (https://hear2read.org/solution-android.php). Only first 7 lessons have been incorporated into the app yet. UI needs to be improved as well.

Demo video for this app - https://www.youtube.com/watch?v=rnR9jGvMHa4

APK and source code for this app is available at - https://github.com/ShivajiDala/SAMVYASA.git

Thanks,

Raghav

To view this discussion on the web visit https://groups.google.com/d/msgid/sanskrit-programmers/CADSGPzV9aRWEB2eT80FmwJDhbXLEPdM0j92sP%3DHsJ-7PXHWX1g%40mail.gmail.com.

[विश्वासो वासुकिजः (Vishvas Vasuki)]

Very interesting! Clicking a few links got met to:

• Andrew Wilkinson, Alok Parlikar, Sunayana Sitaram, Tim White, Alan W Black, Suresh Bazaj, "Open-Source Consumer-Grade Indic Text To Speech", Speech Synthesis Workshop 9 (2016), Sunnyvale, USA. (pdf)

Thence, it looks like festival lite got some Indic language support (though the Sanskrit voice doesn't seem to exist there) - https://github.com/festvox/flite

To view this discussion on the web visit https://groups.google.com/d/msgid/sanskrit-programmers/3d265fbe-de5b-b569-3a44-65621260b56a%40gmail.com.

--

--
Vishvas /विश्वासः

[Irene Galstian]

I’ve never used Sanskrit TTS. Can it be given accented text and asked to reproduce it vocally? If not, how does it determine the intonation with which to render sentences?

Sorry for the basic questions, this is an unfamiliar area. 

On 7 Mar 2021, at 11:27 am, विश्वासो वासुकिजः <vishvas...@gmail.com> wrote:



To view this discussion on the web visit https://groups.google.com/d/msgid/sanskrit-programmers/CAFY6qgFK2c6c7BwNw4OJvRM6dWoLZH0udPMVHq5ewFcfBxRKuw%40mail.gmail.com.

[विश्वासो वासुकिजः (Vishvas Vasuki)]

On Sun, Mar 7, 2021 at 6:07 PM Irene Galstian <gnos...@gmail.com> wrote:

I’ve never used Sanskrit TTS. Can it be given accented text and asked to reproduce it vocally?

If you mean svaras of sanskrit, no - atleast not directly. Maybe someone here knows if some conversion is possible? 

If not, how does it determine the intonation with which to render sentences?

It doesn't - it all sounds pretty monotone if you listen to the samples. (The same defect applies, to a lesser extant, to modern novice sanskrit speakers though.)

 

To view this discussion on the web visit https://groups.google.com/d/msgid/sanskrit-programmers/680226B2-CE5F-46D4-9EE4-93A7DCE9EF1A%40gmail.com.

[Irene Galstian]

Thank you. 

On 7 Mar 2021, at 4:49 pm, विश्वासो वासुकिजः <vishvas...@gmail.com> wrote:



To view this discussion on the web visit https://groups.google.com/d/msgid/sanskrit-programmers/CAFY6qgGnowjf3X_VpPy-qtS2J32Z8qsPawCKQdMEyRiAUQMRGA%40mail.gmail.com.

[Avinash L Varna]

Thanks everyone for the feedback, suggestions and resources!

@Ganesh, Great suggestions! I think it should be possible to use other Indian languages for testing. I'll make some updates to the library to make it easier to choose the intermediate language. I'll also call you sometime this week to discuss ideas and next steps.

@Raghav, This is a great start for the app! I probably can't help much on the UI side. On the data side though, does the app require the sentences to be entered manually, or can it read from a csv, for instance? I see that the sentences are available at https://sa.wiktionary.org/s/5o (Disclaimer - I don't know the license, copyright, etc.). It should be easy to write a script to convert that data to a csv file and load it into the app. For the TTS part, you could use any of the options discussed here to pre-generate the speech for those ~1000 sentences and save it with the app, if needed, or continue to use the hear2read back end.

@Vishvas, Thanks for the paper and links. I don't see any mention of Sanskrit in the papers and documentation. The hear2read website credits the same person for Kannada and Sanskrit voices. I wonder if they are using a similar idea to generate Sanskrit speech via Kannada.

@Irene, I suppose that with a large enough dataset or improved techniques, it might be possible to have a TTS system that understands Vedic intonations sometime in the future. We appear to be far from it today.

Regards,
Avinash

To view this discussion on the web visit https://groups.google.com/d/msgid/sanskrit-programmers/6607E9F7-2951-443C-A2C9-CBDD0A532DE5%40gmail.com.

[Shree Nadig]

Since my last post about developing an ASR system for Sanskrit, we are collaborating with Aksharam, Samskrita Bharati in releasing around 30 hours of clean, read-speech with transcripts and recipes for Kaldi, ESPnet, and pre-trained models.

We decided not to go with the AIR dataset due to uncertainties in licensing rights.

The process of cleaning and aligning the data has started recently.

As I mentioned in my previous post, the intention of this project is to open-source the data, recipe, and models with an easily accessible license to everyone.

We will upload the resources in the near future. You can include that along with the TTS data to improve the ASR models.

We could build a TTS model from the 30-hour data as well, but it will sound robotic as it's multi-speaker data.

There are also talks of using something like mimic-recording-studio to crowd-source the data after the initial 30 hours. This tool could be used to collect data for TTS as well.

Let me know if I can help with anything. I have some experience with Kaldi, ESPnet and other speech toolkits.

----------------------------------------------------

Best Regards,

Shreekantha A Nadig

To view this discussion on the web visit https://groups.google.com/d/msgid/sanskrit-programmers/CAALtx9a-TmrTcaEnv%2BK%3D97x21G3q%2Bp-%3Du98cgMNJCeNuiZuLvQ%40mail.gmail.com.

[Harsh Thakkar]

सर्वेभ्यो नमः  

I have not grasped everything on this thread yet ... but sounds very exciting developments .. 

I thought of chipping in when I saw mention of collaborating with SB (aksharam) as well as crowd-source .. not sure if all are aware - there are enthusiastic volunteers - many with background of both - Samskritam as well as IT .. some of them are eager to contribute to #codeSanskrit .. but they might be very helpful/productive to such initiatives (building corpus/etc) if they get good guidance/direction .. 

[thanks to SB course online at IIT-R which had some 4-5k followers .. that's where they got mobilized from]

this group - which is in formative stage - has started meeting online fortnightly .. their effort to launch official 'opencollective' project has got some 45 stars --> https://github.com/Samskrita-Bharati/zat.am 

shall we join online conf. call to continue some points discussed over this email thread? 

harsh

co...@SamskritaBharati.ca (SB Canada)

जयतु संस्कृतम्

[Gopinath, Raghavendra]

@Avinash, I have collected the sentences using OCR from the SB e-book. One of the reasons I have not put them in a CSV is copyright (it is easy to download; safer tucked away in code).

dhanyavAdAH

Raghav

To view this discussion on the web visit https://groups.google.com/d/msgid/sanskrit-programmers/c2acc066-64eb-41c8-bbe4-cca1ca993ed8n%40googlegroups.com.

[Vinodh Rajan]

Probably it's totally random.

I am trying to map Indic scripts to Hebrew (with vowel diacritics) and I wanted to check how well the mapping worked with the help of a Hebrew TTS by rendering some Sanskrit text in Hebrew.

It doesn't sound half bad :)

https://app.almareader.com/

אֻשְׁנִישַׁ וִגַ׳יַ דָרַנִי . אוֹמְ נַמוֹ בַּגַוַטֵ סַרְוַטְרַַיְלוֹכְַּיַ פְּרַטִוִשִׁשְׁטָיַ בֻּדְדָיַ טֵ נַמַהְ .  טַדְיַטָ .  אוֹמְ בְּרוּמְ בְּרוּמְ בְּרוּמְ .  שׁוֹדַיַ שׁוֹדַיַ . וִשׁוֹדַיַ וִשׁוֹדַיַ .  אַסַמַ סַמַנְטַ אַוַבָּסַסְפַּרַנַ גַטִגַגַנַ סְוַבָּוַ וִשֻׁדְדֵ .  אַבִּשִׁנְצַ׳נְטֻ מָמְ .  סַרְוַטַטָגַטַ סֻגַטַ וַרַוַצַ׳נַ אַמְרֻטַ אַבִּשֵׁכַַּיְהְ .  מַהָמֻדְרַ מַנְטְרַפַּדַַיְהְ .  אָהַרַ אָהַרַ .  מַמַ אָיֻהְ סַנְדָרַנִ .  שׁוֹדַיַ שׁוֹדַיַ . וִשׁוֹדַיַ וִשׁוֹדַיַ .  גַגַנַ סְוַבָּוַ וִשֻׁדְדֵ .  אֻשְׁנִישַׁ וִגַ׳יַ פַּרִשֻׁדְדֵ .  סַהַסְרַ רַשְׁמִ סַמְצ׳וֹדִטֵ .  סַרְוַטַטָגַטַ אַוַלוֹכִַּנִ .  שַׁטְפָּרַמִטָ פַּרִפּוּרַנִ .  סַרְוַטַטָגַטַ מַטִ דַשַׁבּוּמִ פְּרַטִשְׁטִטֵ .  סַרְוַטַטָגַטַ הְרֻדַיַ אַדִשְׁטָנַ אַדִשְׁטִטֵ .  מֻדְרֵ מֻדְרֵ . מַהָמֻדְרֵ .  וַגְ׳רַכַָּיַ סַמְהַרַנַ פַּרִשֻׁדְדֵ .  סַרְוַכַּרְמַ אָוַרַנַ וִשֻׁדְדֵ .  פְּרַטִנִוַרְטַיַ .  מַמַ אָיֻרְוִשֻׁדְדֵ .  סַרְוַטַטָגַטַ סַמַיַ אַדִשְׁטָנַ אַדִשְׁטִטֵ .  אוֹמְ מֻנִ מֻנִ. מַהָמֻנִ וִמֻנִ וִמֻנִ . מַהָוִמֻנִ .  מַטִ מַטִ . מַהָמַטִ . מַמַטִ . סֻמַטִ.  טַטָטָ בּוּטַכַּוֹטִ פַּרִשֻׁדְדֵ .  וִסְפֻּטַ בֻּדְדֵ שֻׁדְדֵ .  הֵ הֵ . גַ׳יַ גַ׳יַ . וִגַ׳יַ וִגַ׳יַ .  סְמַרַ סְמַרַ . סְפַּרַ סְפַּרַ . סְפָּרַיַ סְפָּרַיַ .  סַרְוַבֻּדְדַ אַדִשְׁטָנַ אַדִשְׁטִטֵ .  שֻׁדְדֵ שֻׁדְדֵ . בֻּדְדֵ בֻּדְדֵ .  וַגְ׳רֵ וַגְ׳רֵ . מַהָוַגְ׳רֵ סֻוַגְ׳רֵ .  וַגְ׳רַ גַרְבֵּ . גַ׳יַ גַרְבֵּ . וִגַ׳יַ גַרְבֵּ . וַגְ׳רַגְ׳וָלַ גַרְבֵּ .  וַגְ׳רֵ אֻדְבַּוֵ . וַגְ׳רַ סַמְבַּוֵ וַגְ׳רֵ וַגְ׳רִנִ .  וַגְ׳רַמְ בַּוַטֻ מַמַ שַׁרִירַמְ סַרְוַ סַטְטְוָנָמְ צַ׳ כַָּיַ פַּרִשֻׁדְדִרְבַּוַטֻ .  מֵ סַדָ סַרְוַגַטִ פַּרִשֻׁדְדִשְׁצַ׳ . סַרְוַטַטָגַטַשְׁצַ׳ מָמְ סַמָשְׁוָסַיַנְטֻ .  בֻּדְדְיַ בֻּדְדְיַ . סִדְדְיַ סִדְדְיַ .  בּוֹדַיַ בּוֹדַיַ . וִבּוֹדַיַ וִבּוֹדַיַ .  מוֹצַ׳יַ מוֹצַ׳יַ . וִמוֹצַ׳יַ וִמוֹצַ׳יַ .  שׁוֹדַיַ שׁוֹדַיַ . וִשׁוֹדַיַ וִשׁוֹדַיַ .  סַמַנְטָנַ מוֹצַ׳יַ מוֹצַ׳יַ .  סַמַנְטַ רַשְׁמִ פַּרִשֻׁדְדֵ .  סַרְוַ טַטָגַטַ הְרֻדַיַ אַדִשְׁטָנַ אַדִשְׁטִטֵ .  מֻדְרֵ מֻדְרֵ . מַהָמֻדְרֵ .  מַהָמֻדְרָ מַנְטְרַ פַּדַַיְהְ סְוָהָ ..    אוֹמְ בְּרוּמְ סְוָהָ .  אוֹמְ אַמְרֻטַ אָיֻרְדַדֵ סְוָהָ .

Vinodh

To view this discussion on the web visit https://groups.google.com/d/msgid/sanskrit-programmers/06448d67-4b13-50e0-59ab-670a8ecc0268%40gmail.com.

--

http://www.virtualvinodh.com

[विश्वासो वासुकिजः (Vishvas Vasuki)]

On Sun, Mar 7, 2021 at 10:18 PM विश्वासो वासुकिजः (Vishvas Vasuki) <vishvas...@gmail.com> wrote:

On Sun, Mar 7, 2021 at 6:07 PM Irene Galstian <gnos...@gmail.com> wrote:

I’ve never used Sanskrit TTS. Can it be given accented text and asked to reproduce it vocally?

If you mean svaras of sanskrit, no - atleast not directly. Maybe someone here knows if some conversion is possible? 

उत्कर्षोज्ज्वलाभ्यां (अत्र संयुक्ताभ्यां) राजपुत्राभ्यां सृज्यमानं किञ्चन पुलकयति माम्! तस्य प्राग्रूपम् इह - https://031323.github.io/rgvedopkrnani/path#3365  ("उ॒द्घाट्य॒ कश्चि॑त् का॒लः प्रति॑पाल्यताम्। ततः॒ कामप्यृचं॑ नुदतु। इ॒दानीं॑ प॒ठ्यमा॑नाया ऋ॒चो नोद॑नेन॒ रोद्धुं॑ शक्यते।")

[विश्वासो वासुकिजः (Vishvas Vasuki)]

On Thu, May 20, 2021 at 5:36 PM विश्वासो वासुकिजः (Vishvas Vasuki) <vishvas...@gmail.com> wrote:

On Sun, Mar 7, 2021 at 10:18 PM विश्वासो वासुकिजः (Vishvas Vasuki) <vishvas...@gmail.com> wrote:

On Sun, Mar 7, 2021 at 6:07 PM Irene Galstian <gnos...@gmail.com> wrote:

I’ve never used Sanskrit TTS. Can it be given accented text and asked to reproduce it vocally?

If you mean svaras of sanskrit, no - atleast not directly. Maybe someone here knows if some conversion is possible? 

उत्कर्षोज्ज्वलाभ्यां (अत्र संयुक्ताभ्यां) राजपुत्राभ्यां सृज्यमानं किञ्चन पुलकयति माम्! तस्य प्राग्रूपम् इह - https://031323.github.io/rgvedopkrnani/path#3365  ("उ॒द्घाट्य॒ कश्चि॑त् का॒लः प्रति॑पाल्यताम्। ततः॒ कामप्यृचं॑ नुदतु। इ॒दानीं॑ प॒ठ्यमा॑नाया ऋ॒चो नोद॑नेन॒ रोद्धुं॑ शक्यते।")

एतयोर् https://031323.github.io/suvak/index2.html इत्येतद् अपि मम सौभाग्याय कल्पते!

[विश्वासो वासुकिजः (Vishvas Vasuki)]

On Sun, Mar 7, 2021 at 4:50 AM Avinash L Varna <avinas...@gmail.com> wrote:

Hello all,

I've been investigating cross-lingual TTS techniques to try to adapt available TTS models from other languages for Sanskrit. As part of this investigation, I have been playing around with the Google TTS API and demo - https://cloud.google.com/text-to-speech#section-2.

One experiment I tried is to directly use other Indian languages that mostly have similar phonemes as a substitute for Sanskrit. E.g. use Kannada TTS to generate Sanskrit speech. The results were not bad - definitely intelligible and usable depending on the speed, etc. If you want to try, convert Sanskrit text to Kannada script (e.g. using https://aksharamukha.appspot.com) and paste it in the text section of the demo linked above. Choose Language/locale ಕನ್ನಡ (ಭಾರತ) and one of the available voices. I found that a speed of 0.8 was a good setting, but YMMV.

A way to replicate this experience on a linux desktop environment - 

https://i.imgur.com/aB9XFry.png

 

Since this got tedious after a while, I put together a script to automate it. This weekend, I converted the script into a python package that could be used from other applications - https://github.com/avinashvarna/sanskrit_tts. The package is in a preliminary state, with some defaults that are a personal preference, but could be modified via function arguments. A simple example notebook on using the library is included as well.

For example, I obtained the attached mp3 with the following text from Sanskrit wikipedia (https://sa.wikipedia.org/wiki/कालिदासः#कवेः_देश):

कालिदासस्य जीवनवृत्तिविषये अनेकाः लोकविश्रुतयः अनेके वादाः च सन्ति । केचित् एनं विक्रमादित्यस्य सभायां कविः इति मन्यन्ते । केचित् गुप्तकालीननरेशाणाम् आश्रयं प्राप्तवानिति कथयन्ति । 
धारानगरे भोजराजस्य सभायां कविरत्नपदभूषितः अभूत् इति कथाकोविदाः कथयन्ति। जनश्रुत्यनुसारं बाल्यकाले सः अतीव मूर्खः आसीत् । विद्याधरया सह तस्य विवाहः अभवत् । मूर्खः पतिः इति 
ज्ञात्वा विद्याधरा तं कालीदेव्याः आलयं नीत्वा यावत् सा भवते विद्यां न उपदिशति तावत् भवता ततः बहिः न आगन्तव्यम् इति आदिशत् । ततः पत्न्याः कथनानुसारेण तथैव आचरितः कालिदासः 
कालीदेव्याः वरप्रसादेन विद्वान् अभवत् इति । इयं कथा कालिदासस्य प्रतिभया कविताचातुर्येण च जाता तथ्या तु न इति विदुषां मतिः । यतः जानकीहरणस्य कर्ता राजाकुमारदासः च सप्तमे शतके 
आस्ताम् इति विद्वांसः निश्चितवन्तः । अतः अयं तस्मिन् काले नासीत् इति वक्तुं शक्यते । इतोऽपि कालिदासस्य कविताचातुर्यं निर्दिश्यमानाः अनेकाः कथाः सन्ति ।

There are some systematic errors (e.g. in the pronunciation of visarga following hrasva-i in कविः,पतिः, etc), but the output is probably usable to convert articles/texts to audio for assistive applications, or for listening to articles while commuting/walking/exercising.

Some next steps I am considering:

1. Use this approach to create an initial database of Sanskrit sentences and audio.
   
2. Use database from (1) to create a TTS system that can then be fine-tuned on actual training examples. E.g. maybe fine-tune it on ramayana shloka audio to create a TTS specifically tuned for shlokas as discussed here
3. Over time, add such TTS systems to the sanskrit_tts package
   
4. Maybe use the database from (1) generated with different pitch/speed/etc. settings to bootstrap an ASR system as discussed in the same thread The ASR system could be further fine-tuned on AIR data or others

Any feedback on these steps, and/or the sanskrit_tts library will be appreciated. If you have other ideas for TTS/ASR, please add those as well. Collaborators are always welcome! :)

Regards,

Avinash

--
You received this message because you are subscribed to the Google Groups "sanskrit-programmers" group.
To unsubscribe from this group and stop receiving emails from it, send an email to sanskrit-program...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/sanskrit-programmers/CAALtx9ZiM7i2mYKG0dho5sq16PNoBCOz2sb4bdtGdk7T6-K6jQ%40mail.gmail.com.

--

[विश्वासो वासुकिजः (Vishvas Vasuki)]

I don't know how medium does it, but it's interface (for English) is wonderful. Surprising -

--
You received this message because you are subscribed to the Google Groups "sanskrit-programmers" group.
To unsubscribe from this group and stop receiving emails from it, send an email to sanskrit-program...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/sanskrit-programmers/CAALtx9ZiM7i2mYKG0dho5sq16PNoBCOz2sb4bdtGdk7T6-K6jQ%40mail.gmail.com.