kamathsutra

Actually I noticed that they have some better models now, still narrative in nature, but much better than before.

The whisperspeech one is pretty bad/uncontrollable even in English last I checked, but maybe it can be improved. MahaTTS is quite bad for now. It is trained on IITM data, and that data in my opinion is not great for realistic conversational Hindi TTS. But I need to take a second look.

The way I think about realistic conversational speech is that if you get a phone call, you should not be able to tell whether it is an AI or a human just based on the voice. For English and even some asian languages like Chinese, this has already happened.

If you are a non-Hindi speaker and want to understand the difference, then I might find it difficult to explain :P But whatever you are learning, if you start practicing with a native speaker, I am sure you will easily surpass the SoTA hindi TTS models.

Non-conversational example: https://www.youtube.com/watch?v=ayYk3XkP0ts&t=22s&ab_channel...

You can list to this and understand easily that its AI generated speech. However, it works very well for dubbing etc.

This is speech to text right?

Microsoft has done a lot of work in Speech and Sarvam is also now going into it. But Sarvam is yet to produce a decent LLM for Hindi and now on top of that, they are venturing into speech, so I am not sure how that will pan out. My guess is Sarvam would be more focused on increasing coverage of text to speech for all Indian languages than to make the model indistinguishable from humans for top 4-5 Indian languages. They have some tie-ups with the government for the same.

It's quite robotic.