Upload full repo excluding dump_40, dump_100, precomputed_tokens, precomputed_data

fsq_flow_convnext_training/latest_model.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c497711e5a4ede9155d5b6bebc3cdef0e6a6a5f24bf125219b96cba3c84e30fb
+oid sha256:db74ec68cd52ac3c78172e3db827eeffff7107d2c927e727fa1bff88bd9faeae
 size 663062442

fsq_flow_convnext_training/model_epoch_24.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9b118535b01d1bcf14b0bf18996e77b9efe7bc191d5de6c185f346cd681daacb
+size 663066616

infer_fsq_flow.py

@@ -86,7 +86,7 @@ def phonemize_text(text: str, language_code: str = "de") -> str:
         with_stress=True,
         separator=sep,
         strip=True,
-        preserve_punctuation=False,
+        preserve_punctuation=True,
         njobs=1,
     )
     return ipa

infer_outputs/story.wav

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:518f1591c99cdee38cea5de63f8287e33cb70aa95731b5b65bf6523e8770b202
-size 10617560
+oid sha256:7060fc1d26b1d64f3b03b5f8e6ca633d4e30b5fda1b5c2b8af10deca85c526b7
+size 10695176

train_fsq_flow_convnext.py

@@ -438,30 +438,8 @@ def main(config_path: str, codec_config_path: str):
             phonemes = tgt_texts
             phoneme_lengths = tgt_lens
 
-            # Scheduled duration sampling: occasionally use predicted durations
-            # rescaled to the GT total so upsampled length == codec_lengths (no mismatch).
-            if dur_sampling_prob > 0.0 and random.random() < dur_sampling_prob:
-                unwrapped = getattr(predictor, 'module', predictor)
-                with torch.no_grad():
-                    _dur_out = unwrapped.infer_durations(
-                        phonemes=phonemes,
-                        phoneme_lengths=phoneme_lengths,
-                        speaker_ids=speakers,
-                        language_ids=langs,
-                        ctx_tokens=ctx_texts,
-                        ctx_lens=ctx_lens,
-                        token_start_indices=token_start_indices,
-                        token_end_indices=token_end_indices,
-                        noise_scale=0.3,  # lower noise than val for stability
-                    )
-                _pred_dur = _dur_out['durations'].float()  # [B, T]
-                # Rescale so each sample's duration sum matches GT total.
-                # This keeps codec_lengths valid and the flow loss well-defined.
-                _gt_total = durations.float().sum(dim=-1, keepdim=True).clamp(min=1)  # [B,1]
-                _pred_total = _pred_dur.sum(dim=-1, keepdim=True).clamp(min=1)       # [B,1]
-
-            else:
-                dur_for_model = durations
+
+            dur_for_model = durations
 
             # Forward pass - pass context to flow predictor for contextual encoding
             output = predictor(