From 13c2af5de44d2ac4cb09677ab4fb07f37ade2b98 Mon Sep 17 00:00:00 2001
From: shixian.shi <shixian.shi@alibaba-inc.com>
Date: 星期三, 06 九月 2023 11:51:46 +0800
Subject: [PATCH] fix empty timestamp list inference

---
 funasr/bin/asr_inference_launch.py |   29 ++++++++++++++++-------------
 1 files changed, 16 insertions(+), 13 deletions(-)

diff --git a/funasr/bin/asr_inference_launch.py b/funasr/bin/asr_inference_launch.py
index dda2879..ffb0b26 100644
--- a/funasr/bin/asr_inference_launch.py
+++ b/funasr/bin/asr_inference_launch.py
@@ -368,7 +368,7 @@
             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]
@@ -409,7 +409,7 @@
                         ibest_writer["rtf"][key] = rtf_cur
 
                     if text is not None:
-                        if use_timestamp and timestamp is not None:
+                        if use_timestamp and timestamp is not None and len(timestamp):
                             postprocessed_result = postprocess_utils.sentence_postprocess(token, timestamp)
                         else:
                             postprocessed_result = postprocess_utils.sentence_postprocess(token)
@@ -421,7 +421,7 @@
                         else:
                             text_postprocessed, word_lists = postprocessed_result[0], postprocessed_result[1]
                         item = {'key': key, 'value': text_postprocessed}
-                        if timestamp_postprocessed != "":
+                        if timestamp_postprocessed != "" or len(timestamp) == 0:
                             item['timestamp'] = timestamp_postprocessed
                         asr_result_list.append(item)
                         finish_count += 1
@@ -563,6 +563,8 @@
         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_threshold_s = kwargs.get("batch_size_token_threshold_s", int(speech2vadsegment.vad_model.vad_opts.max_single_segment_time*0.67/1000)) * 1000
         batch_size_token = kwargs.get("batch_size_token", 6000)
         print("batch_size_token: ", batch_size_token)
 
@@ -645,8 +647,7 @@
             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]) < batch_size_token_threshold_s:
                     continue
                 batch_size_token_ms_cum = 0
                 end_idx = j + 1
@@ -685,7 +686,7 @@
             text, token, token_int = result[0], result[1], result[2]
             time_stamp = result[4] if len(result[4]) > 0 else None
 
-            if use_timestamp and time_stamp is not None:
+            if use_timestamp and time_stamp is not None and len(time_stamp):
                 postprocessed_result = postprocess_utils.sentence_postprocess(token, time_stamp)
             else:
                 postprocessed_result = postprocess_utils.sentence_postprocess(token)
@@ -710,7 +711,7 @@
             item = {'key': key, 'value': text_postprocessed_punc}
             if text_postprocessed != "":
                 item['text_postprocessed'] = text_postprocessed
-            if time_stamp_postprocessed != "":
+            if time_stamp_postprocessed != "" or len(time_stamp) == 0:
                 item['time_stamp'] = time_stamp_postprocessed
 
             item['sentences'] = time_stamp_sentence(punc_id_list, time_stamp_postprocessed, text_postprocessed)
@@ -1289,6 +1290,7 @@
         quantize_dtype: Optional[str] = "float16",
         streaming: Optional[bool] = False,
         simu_streaming: Optional[bool] = False,
+        full_utt: Optional[bool] = False,
         chunk_size: Optional[int] = 16,
         left_context: Optional[int] = 16,
         right_context: Optional[int] = 0,
@@ -1338,7 +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"
@@ -1365,14 +1367,12 @@
         quantize_dtype=quantize_dtype,
         streaming=streaming,
         simu_streaming=simu_streaming,
+        full_utt=full_utt,
         chunk_size=chunk_size,
         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,
@@ -1418,7 +1418,7 @@
                         _end = (i + 1) * speech2text._ctx
 
                         speech2text.streaming_decode(
-                            speech[i * speech2text._ctx: _end], is_final=False
+                            speech[i * speech2text._ctx: _end + speech2text._right_ctx], is_final=False
                         )
 
                     final_hyps = speech2text.streaming_decode(
@@ -1426,6 +1426,8 @@
                     )
                 elif speech2text.simu_streaming:
                     final_hyps = speech2text.simu_streaming_decode(**batch)
+                elif speech2text.full_utt:
+                    final_hyps = speech2text.full_utt_decode(**batch)
                 else:
                     final_hyps = speech2text(**batch)
 
@@ -1814,6 +1816,7 @@
     group.add_argument("--ngram_weight", type=float, default=0.9, help="ngram weight")
     group.add_argument("--streaming", type=str2bool, default=False)
     group.add_argument("--simu_streaming", type=str2bool, default=False)
+    group.add_argument("--full_utt", type=str2bool, default=False)
     group.add_argument("--chunk_size", type=int, default=16)
     group.add_argument("--left_context", type=int, default=16)
     group.add_argument("--right_context", type=int, default=0)

--
Gitblit v1.9.1