From 17e8f5b889be2ad31608b5203dc5fbc5fd5c0f8a Mon Sep 17 00:00:00 2001
From: nichongjia-2007 <nichongjia@gmail.com>
Date: 星期四, 20 七月 2023 21:26:58 +0800
Subject: [PATCH] Merge branch 'main' of https://github.com/alibaba-damo-academy/FunASR

---
 funasr/bin/asr_inference_launch.py |   16 +++++++---------
 1 files changed, 7 insertions(+), 9 deletions(-)

diff --git a/funasr/bin/asr_inference_launch.py b/funasr/bin/asr_inference_launch.py
index 10f8e50..36c6d76 100644
--- a/funasr/bin/asr_inference_launch.py
+++ b/funasr/bin/asr_inference_launch.py
@@ -370,7 +370,7 @@
             results = speech2text(**batch)
             if len(results) < 1:
                 hyp = Hypothesis(score=0.0, scores={}, states={}, yseq=[])
-                results = [[" ", ["sil"], [2], hyp, 10, 6]] * nbest
+                results = [[" ", ["sil"], [2], hyp, 10, 6, []]] * nbest
             time_end = time.time()
             forward_time = time_end - time_beg
             lfr_factor = results[0][-1]
@@ -439,6 +439,7 @@
         logging.info(rtf_avg)
         if writer is not None:
             ibest_writer["rtf"]["rtf_avf"] = rtf_avg
+        torch.cuda.empty_cache()
         return asr_result_list
 
     return _forward
@@ -564,6 +565,7 @@
         if 'hotword' in kwargs:
             hotword_list_or_file = kwargs['hotword']
 
+        speech2vadsegment.vad_model.vad_opts.max_single_segment_time = kwargs.get("max_single_segment_time", 60000)
         batch_size_token = kwargs.get("batch_size_token", 6000)
         print("batch_size_token: ", batch_size_token)
 
@@ -646,8 +648,7 @@
             beg_idx = 0
             for j, _ in enumerate(range(0, n)):
                 batch_size_token_ms_cum += (sorted_data[j][0][1] - sorted_data[j][0][0])
-                if j < n - 1 and (batch_size_token_ms_cum + sorted_data[j + 1][0][1] - sorted_data[j + 1][0][
-                    0]) < batch_size_token_ms:
+                if j < n - 1 and (batch_size_token_ms_cum + sorted_data[j + 1][0][1] - sorted_data[j + 1][0][0]) < batch_size_token_ms and (sorted_data[j + 1][0][1] - sorted_data[j + 1][0][0]) < speech2vadsegment.vad_model.vad_opts.max_single_segment_time:
                     continue
                 batch_size_token_ms_cum = 0
                 end_idx = j + 1
@@ -730,6 +731,7 @@
                     ibest_writer["time_stamp"][key] = "{}".format(time_stamp_postprocessed)
 
             logging.info("decoding, utt: {}, predictions: {}".format(key, text_postprocessed_punc))
+        torch.cuda.empty_cache()
         return asr_result_list
 
     return _forward
@@ -1327,7 +1329,6 @@
         right_context: Number of frames in right context AFTER subsampling.
         display_partial_hypotheses: Whether to display partial hypotheses.
     """
-    # assert check_argument_types()
 
     if batch_size > 1:
         raise NotImplementedError("batch decoding is not implemented")
@@ -1339,7 +1340,7 @@
         format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s",
     )
 
-    if ngpu >= 1:
+    if ngpu >= 1 and torch.cuda.is_available():
         device = "cuda"
     else:
         device = "cpu"
@@ -1370,10 +1371,7 @@
         left_context=left_context,
         right_context=right_context,
     )
-    speech2text = Speech2TextTransducer.from_pretrained(
-        model_tag=model_tag,
-        **speech2text_kwargs,
-    )
+    speech2text = Speech2TextTransducer(**speech2text_kwargs)
 
     def _forward(data_path_and_name_and_type,
                  raw_inputs: Union[np.ndarray, torch.Tensor] = None,

--
Gitblit v1.9.1