How can we have different silence?

[cod3r0k]

Dear @shivammehta25
We'd like system allows different pause/silence durations after specific characters. For example:

• A comma (,) might have a 0.2s pause
• A semicolon (;) might need a slightly longer pause (e.g., 0.35s)
• Custom bullet characters (like • or –) could have even longer pauses

This is especially useful when generating more natural-sounding speech or differentiating between sentence-level and list-level pauses.

Use Cases:
Control pause after custom punctuation
Fine-tune prosody for lists, bullet points, and long-form content
Better expression in multilingual or formal text structures

Suggestions:
Allow configuration of pause duration per character via a mapping table
Extend pre-processing to insert appropriate SSML tags
Or, support a post-processing layer that adds tags before feeding to TTS engine

Can you help me?

[cod3r0k]

Or for example add some option in text like SIL, PAUSE, BREAK or manually say SIL1000 for 1 sec or SIL5000 for 5 sec @shivammehta25

[cod3r0k]

Hi @shivammehta25 do you have any guide about that?

[cod3r0k]

Any update? @shivammehta25

[cod3r0k]

Plz @shivammehta25

[shivammehta25]

Hi @cod3r0k

Apologies for the delayed response, but I was away. Since this is a research project, we currently do not plan to support it beyond what was presented in the research paper.

However, the code is under MIT license, you can fork it and modify it as you want. I will suggest that all things you've mentioned would require annotations to the dataset. Once you have your dataset annotated you can change the text processing to deal with it in a very similar manner as a new language or increasing the vocabulary for the current language.

Let me know if you have any further questions.

Regards,
Shivam

[cod3r0k]

Hi @cod3r0k

Apologies for the delayed response, but I was away. Since this is a research project, we currently do not plan to support it beyond what was presented in the research paper.

However, the code is under MIT license, you can fork it and modify it as you want. I will suggest that all things you've mentioned would require annotations to the dataset. Once you have your dataset annotated you can change the text processing to deal with it in a very similar manner as a new language or increasing the vocabulary for the current language.

Let me know if you have any further questions.

Regards, Shivam

Dear @shivammehta25
Thanks for your response and for sharing the paper—it’s been very helpful.

As a follow-up, I was wondering if you might have any suggestions on how to approach this idea from a more scientific or architectural perspective?

For example, let’s say I want to implement 2 seconds of silence after a period (.) and 5 seconds after a colon (:)—how would you recommend approaching this? Should it be handled in:

Text preprocessing (e.g., inserting SSML <break> tags),

Dataset annotation (e.g., marking these punctuation-based pauses during training),

Or modifying the model’s prosody conditioning (if supported)?

I’d really appreciate any ideas or references you might have on how to systematically and efficiently add this kind of fine-grained pause control—especially for long-form or formal TTS applications.

Looking forward to your insights.

[cod3r0k]

Dear @shivammehta25 I want pause what should I do ? for example add | for more pause.

[shivammehta25]

I think it should be a combination of both, you should have dataset annotation and handle these pauses in text preprocessing, and let the model learn those. Currently, LJ Speech and VCTK, which are read speech (and, quite frankly, have a bit weird audio), lack good pauses in the dataset. You will need a better dataset, which you can annotate and train on then you can expect the pauses to be natural.