From 17d9821922e1024de46063d3de57c5acf5a2a4e1 Mon Sep 17 00:00:00 2001
From: KakaruHayate <kakaru.hk@hotmail.com>
Date: Wed, 15 Jul 2026 14:42:40 +0800
Subject: [PATCH 1/2] fix: pad mel with spec_min instead of 0.0 in acoustic
 collater
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Raw log-mel 0.0 is not a neutral padding value: norm_spec maps it to
+1.0 — the very top of the normalized range, i.e. maximum loudness.
Every batch therefore filled the padded tail of shorter samples with
full-loudness garbage. Two consequences:

1. The diffusion backbones (WaveNet / LYNXNet / LYNXNet2) receive no
   padding mask, so their receptive field (~181 frames for LYNXNet2
   with kernel_size=31 x 6 layers) leaks the fake signal into the
   trailing valid frames. The loss mask (mel2ph > 0) hides this from
   the loss on padding frames, but the contaminated valid frames near
   the boundary are fully counted — a systematic bias on utterance
   tails, exactly where breathy endings and vibrato decay live.

2. The aux decoder loss is not masked at all: with zero-padding the
   aux decoder was actively trained to predict maximum loudness on
   padding frames from near-zero condition.

Padding with spec_min (-12 by default) maps to -1.0 (silence) and sits
next to the mel extractor's true silence floor log(1e-5) = -11.51, so
padded regions now look like ordinary trailing silence — consistent
with what the model sees at inference time.

Note: this changes the training data distribution slightly; models
trained before/after this fix are checkpoint-compatible but their
padded-region behavior differs.
---
 training/acoustic_task.py | 12 +++++++++++-
 1 file changed, 11 insertions(+), 1 deletion(-)

diff --git a/training/acoustic_task.py b/training/acoustic_task.py
index ca6a71c65..f1ee4ab91 100644
--- a/training/acoustic_task.py
+++ b/training/acoustic_task.py
@@ -45,7 +45,17 @@ def collater(self, samples):
         tokens = utils.collate_nd([s['tokens'] for s in samples], 0)
         f0 = utils.collate_nd([s['f0'] for s in samples], 0.0)
         mel2ph = utils.collate_nd([s['mel2ph'] for s in samples], 0)
-        mel = utils.collate_nd([s['mel'] for s in samples], 0.0)
+        # Pad mel with spec_min (silence) instead of 0.0. Raw log-mel 0.0 lies at
+        # the TOP of the normalized range (norm_spec maps it to +1.0, i.e. maximum
+        # loudness), so zero-padding used to fill the padded tail of every shorter
+        # sample with full-loudness garbage. The diffusion loss masks these frames,
+        # but the backbone has no padding mask and its receptive field (~181 frames
+        # for LYNXNet2 k=31 x 6 layers) leaks the fake signal into the trailing
+        # valid frames; the aux decoder loss is not masked at all. spec_min (-12)
+        # sits right at the true silence floor log(1e-5) = -11.51 of the mel
+        # extractor, i.e. padding now looks like silence.
+        mel_pad = float(hparams['spec_min'][0]) if hparams.get('spec_min') else -12.0
+        mel = utils.collate_nd([s['mel'] for s in samples], mel_pad)
         batch.update({
             'tokens': tokens,
             'mel2ph': mel2ph,

From 5f2a10fa96d3706006f73f3be458b52fc3fce9af Mon Sep 17 00:00:00 2001
From: Kakaru <97896816+KakaruHayate@users.noreply.github.com>
Date: Wed, 15 Jul 2026 14:51:12 +0800
Subject: [PATCH 2/2] Update mel padding strategy to use spec_min

Replace zero-padding with spec_min for mel padding to avoid full-loudness garbage in shorter samples.
---
 training/acoustic_task.py | 9 ---------
 1 file changed, 9 deletions(-)

diff --git a/training/acoustic_task.py b/training/acoustic_task.py
index f1ee4ab91..a9f514eac 100644
--- a/training/acoustic_task.py
+++ b/training/acoustic_task.py
@@ -45,15 +45,6 @@ def collater(self, samples):
         tokens = utils.collate_nd([s['tokens'] for s in samples], 0)
         f0 = utils.collate_nd([s['f0'] for s in samples], 0.0)
         mel2ph = utils.collate_nd([s['mel2ph'] for s in samples], 0)
-        # Pad mel with spec_min (silence) instead of 0.0. Raw log-mel 0.0 lies at
-        # the TOP of the normalized range (norm_spec maps it to +1.0, i.e. maximum
-        # loudness), so zero-padding used to fill the padded tail of every shorter
-        # sample with full-loudness garbage. The diffusion loss masks these frames,
-        # but the backbone has no padding mask and its receptive field (~181 frames
-        # for LYNXNet2 k=31 x 6 layers) leaks the fake signal into the trailing
-        # valid frames; the aux decoder loss is not masked at all. spec_min (-12)
-        # sits right at the true silence floor log(1e-5) = -11.51 of the mel
-        # extractor, i.e. padding now looks like silence.
         mel_pad = float(hparams['spec_min'][0]) if hparams.get('spec_min') else -12.0
         mel = utils.collate_nd([s['mel'] for s in samples], mel_pad)
         batch.update({