fix: SenseVoice+spk_model crash and torchaudio>=2.11 compatibility

1. auto_model.py: When ASR model (e.g. SenseVoice) doesn't produce
   timestamps, automatically fall back to vad_segment mode for speaker
   diarization instead of crashing. Fixes #2945.

2. load_utils.py: Add soundfile as fallback when torchaudio.load fails
   (torchaudio>=2.11 requires torchcodec). Falls through to ffmpeg only
   if soundfile also fails.

funasr/auto/auto_model.py

@@ -819,6 +819,9 @@ class AutoModel:
                 )
                 # del result['spk_embedding']
                 sv_output = postprocess(all_segments, None, labels, spk_embedding.cpu())
+                if self.spk_mode == "punc_segment" and "timestamp" not in result and "timestamps" not in result:
+                    logging.warning("No timestamps in ASR result (e.g. SenseVoice), falling back to vad_segment mode for speaker diarization.")
+                    self.spk_mode = "vad_segment"
                 if self.spk_mode == "vad_segment":  # recover sentence_list
                     sentence_list = []
                     for rest, vadsegment in zip(restored_data, vadsegments):

funasr/utils/load_utils.py

@@ -113,10 +113,17 @@ def load_audio_text_image_video(
                 if kwargs.get("reduce_channels", True):
                     data_or_path_or_list = data_or_path_or_list.mean(0)
             except:
-                data_or_path_or_list = _load_audio_ffmpeg(data_or_path_or_list, sr=fs)
-                data_or_path_or_list = torch.from_numpy(
-                    data_or_path_or_list
-                ).squeeze()  # [n_samples,]
+                try:
+                    import soundfile as sf
+                    data_np, audio_fs = sf.read(data_or_path_or_list, dtype="float32")
+                    data_or_path_or_list = torch.from_numpy(data_np).squeeze()
+                    if data_or_path_or_list.ndim > 1 and kwargs.get("reduce_channels", True):
+                        data_or_path_or_list = data_or_path_or_list.mean(-1)
+                except:
+                    data_or_path_or_list = _load_audio_ffmpeg(data_or_path_or_list, sr=fs)
+                    data_or_path_or_list = torch.from_numpy(
+                        data_or_path_or_list
+                    ).squeeze()  # [n_samples,]
         elif data_type == "text" and tokenizer is not None:
             with open(data_or_path_or_list, "r") as f:
                 data_or_path_or_list = tokenizer.encode(f.read().strip())