Merge pull request #441 from alibaba-damo-academy/dev_lhn

update

funasr/bin/asr_inference_paraformer_streaming.py

@@ -205,9 +205,12 @@ class Speech2Text:
         results = []
         cache_en = cache["encoder"]
         if speech.shape[1] < 16 * 60 and cache_en["is_final"]:
+            if cache_en["start_idx"] == 0:
+                return []
             cache_en["tail_chunk"] = True
             feats = cache_en["feats"]
             feats_len = torch.tensor([feats.shape[1]])
+            self.asr_model.frontend = None
             results = self.infer(feats, feats_len, cache)
             return results
         else:

funasr/models/encoder/sanm_encoder.py

@@ -380,7 +380,7 @@ class SANMEncoder(AbsEncoder):
         else:
             xs_pad = self.embed(xs_pad, cache)
         if cache["tail_chunk"]:
-            xs_pad = cache["feats"]
+            xs_pad = to_device(cache["feats"], device=xs_pad.device)
         else:
             xs_pad = self._add_overlap_chunk(xs_pad, cache)
         encoder_outs = self.encoders0(xs_pad, None, None, None, None)