fix: prevent sample loss in chunked_decode when decoder returns fewer frames

finetuning/dataset.py

@@ -102,7 +102,9 @@ def _tokenize_texts(self, text) -> List[torch.Tensor]:
 
     @torch.inference_mode()
     def extract_mels(self, audio, sr):
-        assert sr == 24000, "Only support 24kHz audio"
+        if sr != 24000:
+            audio = librosa.resample(audio, orig_sr=sr, target_sr=24000)
+            sr = 24000
         mels = mel_spectrogram(
             torch.from_numpy(audio).unsqueeze(0), 
             n_fft=1024, 

finetuning/sft_12hz.py

@@ -147,10 +147,8 @@ def train():
             unwrapped_model = accelerator.unwrap_model(model)
             state_dict = {k: v.detach().to("cpu") for k, v in unwrapped_model.state_dict().items()}
 
-            drop_prefix = "speaker_encoder"
-            keys_to_drop = [k for k in state_dict.keys() if k.startswith(drop_prefix)]
-            for k in keys_to_drop:
-                del state_dict[k]
+            # NOTE: speaker_encoder weights are preserved in checkpoints to allow
+            # training resume. They can be stripped for inference-only exports.
 
             weight = state_dict['talker.model.codec_embedding.weight']
             state_dict['talker.model.codec_embedding.weight'][3000] = target_speaker_embedding[0].detach().to(weight.device).to(weight.dtype)

qwen_tts/core/tokenizer_12hz/modeling_qwen3_tts_tokenizer_v2.py

@@ -891,7 +891,8 @@ def chunked_decode(self, codes, chunk_size=300, left_context_size=25):
             context_size = left_context_size if start_index - left_context_size > 0 else start_index
             codes_chunk = codes[..., start_index - context_size : end_index]
             wav_chunk = self(codes_chunk)
-            wavs.append(wav_chunk[..., context_size * self.total_upsample :])
+            sample_count = min((end_index - start_index) * self.total_upsample, wav_chunk.shape[-1])
+            wavs.append(wav_chunk[..., -sample_count:])
             start_index = end_index
         return torch.cat(wavs, dim=-1)