fix(omnivoice): pass duration to model.generate instead of truncating

[Arturogpj]

Problem

When a user lowers the duration slider for OmniVoice, WanGP returns a slice of the end of a naturally-paced generation instead of compressing the speech to fit the requested length.

Root cause

OmniVoice's model.generate(duration=X) is the upstream-supported way to fix output length (see k2-fsa/OmniVoice — "duration: fixed output duration in seconds. Overrides speed when set."). It sets target_lens = X * frame_rate and computes a speed ratio so the model fits the text into that duration.

WanGP's OmniVoicePipeline never passed duration= to model.generate() and instead generated at the model's natural pace and truncated the tail afterwards (pipeline.py _generate_segments):

audios = self.model.generate(text=..., generation_config=...)  # no duration=
...
segment_audio = segment_audio[:samples_left]  # chop off whatever didn't fit

So lowering the duration just cut off whatever the model hadn't had time to produce — not compression.

Fix

In app/models/TTS/omnivoice/pipeline.py:

1. _run_segment now accepts a segment_duration param and passes it to self.model.generate(duration=segment_duration). The model now fits each segment's text into its allotted seconds — true compression via the upstream API.

2. _generate_segments now distributes the requested duration across segments proportionally to their estimated speech seconds (using the existing _estimate_text_seconds helper). Inter-segment pauses count toward the request, so the final file length ≈ the requested duration. The old post-hoc truncation is replaced with a small safety trim (a few hundred ms) that only guards against post-processing silence overshoot — never a content cut.

3. Auto mode (duration_seconds=0/None) is unchanged — segment_duration=None is passed to every segment, preserving the previous behavior.

Verification

Tested locally via the Wan2GP UI:

• Same text at duration=15s → full-length natural speech (unchanged).
• Same text at duration=5s → full text compressed (faster speech), not a 5s slice of the 15s output. ✅
• duration=0 (auto) → unchanged behavior. ✅

Scope

Single file, ~36 lines added, 2 removed. No changes to modeling_omnivoice.py, the handler, or any other TTS pipeline.