Merge pull request #2000 from yt605155624/fix_chunk

[TTS]reneame chunk to block in streaming tts, test=tts

paddlespeech/t2s/exps/inference_streaming.py

@@ -90,7 +90,7 @@ def parse_args():
         default=False,
         help="whether use streaming acoustic model")
     parser.add_argument(
-        "--chunk_size", type=int, default=42, help="chunk size of am streaming")
+        "--block_size", type=int, default=42, help="block size of am streaming")
     parser.add_argument(
         "--pad_size", type=int, default=12, help="pad size of am streaming")
 
@@ -169,7 +169,7 @@ def main():
 
     N = 0
     T = 0
-    chunk_size = args.chunk_size
+    block_size = args.block_size
     pad_size = args.pad_size
     get_tone_ids = False
     for utt_id, sentence in sentences:
@@ -189,7 +189,7 @@ def main():
                 am_encoder_infer_predictor, input=phones)
 
             if args.am_streaming:
-                hss = get_chunks(orig_hs, chunk_size, pad_size)
+                hss = get_chunks(orig_hs, block_size, pad_size)
                 chunk_num = len(hss)
                 mel_list = []
                 for i, hs in enumerate(hss):
@@ -211,7 +211,7 @@ def main():
                         sub_mel = sub_mel[pad_size:]
                     else:
                         # 倒数几块的右侧也可能没有 pad 够
-                        sub_mel = sub_mel[pad_size:(chunk_size + pad_size) -
+                        sub_mel = sub_mel[pad_size:(block_size + pad_size) -
                                           sub_mel.shape[0]]
                     mel_list.append(sub_mel)
                 mel = np.concatenate(mel_list, axis=0)

paddlespeech/t2s/exps/ort_predict_streaming.py

@@ -97,7 +97,7 @@ def ort_predict(args):
     T = 0
     merge_sentences = True
     get_tone_ids = False
-    chunk_size = args.chunk_size
+    block_size = args.block_size
     pad_size = args.pad_size
 
     for utt_id, sentence in sentences:
@@ -115,7 +115,7 @@ def ort_predict(args):
             orig_hs = am_encoder_infer_sess.run(
                 None, input_feed={'text': phone_ids})
             if args.am_streaming:
-                hss = get_chunks(orig_hs[0], chunk_size, pad_size)
+                hss = get_chunks(orig_hs[0], block_size, pad_size)
                 chunk_num = len(hss)
                 mel_list = []
                 for i, hs in enumerate(hss):
@@ -139,7 +139,7 @@ def ort_predict(args):
                         sub_mel = sub_mel[pad_size:]
                     else:
                         # 倒数几块的右侧也可能没有 pad 够
-                        sub_mel = sub_mel[pad_size:(chunk_size + pad_size) -
+                        sub_mel = sub_mel[pad_size:(block_size + pad_size) -
                                           sub_mel.shape[0]]
                     mel_list.append(sub_mel)
                 mel = np.concatenate(mel_list, axis=0)
@@ -236,7 +236,7 @@ def parse_args():
         default=False,
         help="whether use streaming acoustic model")
     parser.add_argument(
-        "--chunk_size", type=int, default=42, help="chunk size of am streaming")
+        "--block_size", type=int, default=42, help="block size of am streaming")
     parser.add_argument(
         "--pad_size", type=int, default=12, help="pad size of am streaming")
 

paddlespeech/t2s/exps/syn_utils.py

@@ -75,13 +75,13 @@ def denorm(data, mean, std):
     return data * std + mean
 
 
-def get_chunks(data, chunk_size: int, pad_size: int):
+def get_chunks(data, block_size: int, pad_size: int):
     data_len = data.shape[1]
     chunks = []
-    n = math.ceil(data_len / chunk_size)
+    n = math.ceil(data_len / block_size)
     for i in range(n):
-        start = max(0, i * chunk_size - pad_size)
+        start = max(0, i * block_size - pad_size)
-        end = min((i + 1) * chunk_size + pad_size, data_len)
+        end = min((i + 1) * block_size + pad_size, data_len)
         chunks.append(data[:, start:end, :])
     return chunks
 

paddlespeech/t2s/exps/synthesize_streaming.py

@@ -133,7 +133,7 @@ def evaluate(args):
 
     N = 0
     T = 0
-    chunk_size = args.chunk_size
+    block_size = args.block_size
     pad_size = args.pad_size
 
     for utt_id, sentence in sentences:
@@ -153,7 +153,7 @@ def evaluate(args):
                 # acoustic model
                 orig_hs = am_encoder_infer(phone_ids)
                 if args.am_streaming:
-                    hss = get_chunks(orig_hs, chunk_size, pad_size)
+                    hss = get_chunks(orig_hs, block_size, pad_size)
                     chunk_num = len(hss)
                     mel_list = []
                     for i, hs in enumerate(hss):
@@ -171,7 +171,7 @@ def evaluate(args):
                             sub_mel = sub_mel[pad_size:]
                         else:
                             # 倒数几块的右侧也可能没有 pad 够
-                            sub_mel = sub_mel[pad_size:(chunk_size + pad_size) -
+                            sub_mel = sub_mel[pad_size:(block_size + pad_size) -
                                               sub_mel.shape[0]]
                         mel_list.append(sub_mel)
                     mel = paddle.concat(mel_list, axis=0)
@@ -277,7 +277,7 @@ def parse_args():
         default=False,
         help="whether use streaming acoustic model")
     parser.add_argument(
-        "--chunk_size", type=int, default=42, help="chunk size of am streaming")
+        "--block_size", type=int, default=42, help="block size of am streaming")
     parser.add_argument(
         "--pad_size", type=int, default=12, help="pad size of am streaming")