Did you Simplify the Emotion Control?

[frink]

IndexTTS2 has 3 types of emotion control:

1. Text
2. Sliders
3. Reference

If this was tampered with it could explain the difference in encoding.

Otherwise, not sure.

[frink]

Wait could your FFT be different? That would cause some noise...

[DevMan57]

I'll have another look at this over the weekend, I've been busy on other projects. I have a fresh, more refined approach planned which takes advantage of agent teams. I've tried this setup with another project and it is amazing.

[frink]

Let me know how it goes.

[DevMan57]

IndexTTS2 has 3 types of emotion control:

1. Text
2. Sliders
3. Reference

If this was tampered with it could explain the difference in encoding.

Otherwise, not sure.

Yeah I suspect something like this happened, like the agent assumed the typical architecture as they like to default to the average. Serves me right for not knowing how to code lol.

[DevMan57]

Putting some more work in today, after making some progress recently (I assume, wont know for sure until I hear a clean .wav).

[DevMan57]

Let me know how it goes.

Looks like you were right:

› 1. Fix emotion control (Recommended) Align Rust behavior with Python emotion controls (text/
sliders/reference).
2. Debug core encoding mismatch Ignore emotion features and focus on speaker/semantic
encoding differences.
3. Both in one plan Scope a plan that fixes emotion control and encoding
issues together.
4. None of the above Optionally, add details in notes (tab).

[DevMan57]

IndexTTS2-Rust
Rust port of IndexTTS2 with CUDA inference, voice cloning, and emotion control.

Status (February 11, 2026) DONT LISTEN TO CODEX, YES IT PRODUCES AN ACTUAL VOICE NOW BUT CLONING DOESN'T WORK, QUALITY NEEDS IMPROVING, SOMETHING IS OFF WITH EMOTIONS...
The pipeline is now producing intelligible speech on GPU. The major "rumbling water" failure mode was fixed.

Key fixes completed
GPT logits path parity fix:
Added separate LM-head norm loading from final_norm.* and applied it before mel projection.
Multi-step GPT parity instrumentation:
Added step1/last-step logits and cache-length dump/compare support.
DiT parity alignment:
Added RMSNorm AdaLN behavior, RoPE, UViT skip schedule fixes, and adaptive final norm behavior.
BigVGAN critical fix:
Replaced approximate upsampling with true ConvTranspose1d in Rust.
Removed extra non-upstream Snake activation before upsample blocks.
Post-vocoder stability guard:
Added de-rumble high-pass option and automatic fallback for high DC-bias outputs.
Current parity/quality snapshot
DiT parity is very close (step-level diffs in low 1e-4 range in latest DiT parity run).
GPT step0 logits are close; step1+ drift remains and is the next parity target.
End-to-end GPU inference now outputs audible speech files.
What works now
Voice cloning from a reference speaker clip.
Emotion control pathways:
Emotion reference audio (--emotion-audio)
Emotion audio blending (--emotion-audio + --emotion-alpha <0..1>)
Manual emotion vector (--emotion-vector)
Emotion from text (--use-emo-text and optional --emo-text)