Repository navigation
Replies: 2 comments
|
Hi @chucly112, Thank you for the thoughtful suggestion — this is actually a very important direction for making VieNeu-TTS more production-ready. You are absolutely right that style and emotional control (e.g., formal vs casual) are critical for real-world applications such as assistants, narration, and media content. ✅ About Style Control via TokensVieNeu-TTS is based on a language-model style architecture rather than a traditional TTS pipeline. In LM-based TTS systems, style is typically learned as part of the token distribution rather than through a separate module. Adding control tokens allows the model to condition prosody, pacing, and speaking behavior without modifying the core model structure. So yes — prefix style tokens are a promising direction and align well with future extensibility. ✅ Dataset ConsiderationsWhether this works well depends less on the token itself and more on stylistic diversity in the training data. If the dataset contains:
then the model can learn meaningful style separation. Otherwise, style tokens may collapse into nearly identical prosody. This is something we are currently evaluating in the existing 1000h dataset. ✅ Style Tags vs Reference AudioBoth approaches have trade-offs:
A hybrid approach (token + reference guidance) is likely the long-term direction. 🚧 Current Status / Future DirectionExplicit emotion or style conditioning is not yet implemented in the current release, but controllability (style, speaking rate, expressive prosody) is an area we are actively interested in exploring for future versions. Community experimentation in this area is very welcome. Thanks again for the excellent proposal! Best, |
|
Hi @pnnbao97, Thank you for the detailed and insightful response! It's exciting to hear that the LM-based architecture of VieNeu-TTS naturally supports special control tokens. Regarding the stylistic diversity in the 1000h dataset: Also, for the Hybrid approach, do you think adjusting sampling parameters (like Looking forward to seeing how VieNeu-TTS evolves! Best regards, |
Uh oh!
There was an error while loading. Please reload this page.
Hi @pnnbao97 and the VieNeu-TTS team,
First of all, thank you for this incredible project! VieNeu-TTS is already doing an amazing job with Vietnamese naturalness and speed. To take it to the next level for production use (AI Assistants, Podcasts, News), I’d like to suggest adding Style Control Tags.
Based on common use-cases in Vietnam, I propose grouping these styles into two main clusters:
1.
<formal>(Professional/Narrative)2.
<casual>(Conversational/Friendly)Potential Approach
Questions for the team:
I believe these additions would make VieNeu-TTS the most comprehensive and "human-sounding" TTS engine for the Vietnamese language.
Looking forward to hearing your thoughts!
Best regards,
All reactions