Backport inference fixes from upstream PRs QwenLM#234 and QwenLM#126

- QwenLM#234: Fix sample loss in chunked_decode — the context removal logic
  over-trims audio at chunk boundaries, causing shorter output and
  audible pops. Take expected sample count from the end instead of
  trimming a fixed amount from the front.

- QwenLM#126: Fix tuple item assignment in _normalize_audio_inputs — mutating
  a tuple element silently fails, breaking stereo-to-mono conversion.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: guybrush1984

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)
 

qwen_tts/inference/qwen3_tts_model.py

@@ -259,8 +259,8 @@ def _normalize_audio_inputs(self, audios: Union[AudioLike, List[AudioLike]]) ->
                 raise TypeError(f"Unsupported audio input type: {type(a)}")
         for i, a in enumerate(out):
             if a[0].ndim > 1:
-                a[0] = np.mean(a[0], axis=-1).astype(np.float32)
-                out[i] = (a[0], a[1])
+                a_s = np.mean(a[0], axis=-1).astype(np.float32)
+                out[i] = (a_s, a[1])
         return out
 
     def _ensure_list(self, x: MaybeList) -> List[Any]: