Merge branch 'PaddlePaddle:develop' into develop

paddlespeech/cli/whisper/infer.py

@@ -152,8 +152,7 @@ class WhisperExecutor(BaseExecutor):
         Init model and other resources from a specific path.
         """
         logger.debug("start to init the model")
-        # default max_len: unit:second
-        self.max_len = 50
+
         if hasattr(self, 'model'):
             logger.debug('Model had been initialized.')
             return
@@ -339,12 +338,6 @@ class WhisperExecutor(BaseExecutor):
         try:
             audio, audio_sample_rate = soundfile.read(
                 audio_file, dtype="int16", always_2d=True)
-            audio_duration = audio.shape[0] / audio_sample_rate
-            if audio_duration > self.max_len:
-                logger.error(
-                    f"Please input audio file less then {self.max_len} seconds.\n"
-                )
-                return False
         except Exception as e:
             logger.exception(e)
             logger.error(

paddlespeech/s2t/models/whisper/__init__.py

@@ -1,5 +1,5 @@
 # MIT License, Copyright (c) 2022 OpenAI.
-# Copyright (c) 2022 PaddlePaddle Authors and . All Rights Reserved.
+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
 # 
 # Modified from OpenAI Whisper 2022 (https://github.com/openai/whisper/whisper/__init__.py)
 from paddlespeech.s2t.models.whisper.whipser import decode

paddlespeech/s2t/models/whisper/tokenizer.py

@@ -1,5 +1,5 @@
 # MIT License, Copyright (c) 2022 OpenAI.
-# Copyright (c) 2022 PaddlePaddle Authors and . All Rights Reserved.
+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
 # 
 # Modified from OpenAI Whisper 2022 (https://github.com/openai/whisper/whisper/tokenizer.py)
 import os

paddlespeech/s2t/models/whisper/utils.py

@@ -1,5 +1,5 @@
 # MIT License, Copyright (c) 2022 OpenAI.
-# Copyright (c) 2022 PaddlePaddle Authors and . All Rights Reserved.
+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
 # 
 # Modified from OpenAI Whisper 2022 (https://github.com/openai/whisper/whisper/utils.py)
 import zlib

paddlespeech/s2t/models/whisper/whipser.py

@@ -1,5 +1,5 @@
 # MIT License, Copyright (c) 2022 OpenAI.
-# Copyright (c) 2022 PaddlePaddle Authors and . All Rights Reserved.
+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
 #
 # Modified from OpenAI Whisper 2022 (https://github.com/openai/whisper/whisper)
 import os

paddlespeech/t2s/modules/diffusion.py

@@ -360,6 +360,8 @@ class GaussianDiffusion(nn.Layer):
                   num_inference_steps: Optional[int]=1000,
                   strength: Optional[float]=None,
                   scheduler_type: Optional[str]="ddpm",
+                  clip_noise: Optional[bool]=True,
+                  clip_noise_range: Optional[Tuple[float, float]]=(-1, 1),
                   callback: Optional[Callable[[int, int, int, paddle.Tensor],
                                               None]]=None,
                   callback_steps: Optional[int]=1):
@@ -380,6 +382,10 @@ class GaussianDiffusion(nn.Layer):
             scheduler_type (str, optional):
                 Noise scheduler for generate noises. 
                 Choose a great scheduler can skip many denoising step, by default 'ddpm'.
+            clip_noise (bool, optional):
+                Whether to clip each denoised output, by default True.
+            clip_noise_range (tuple, optional):
+                denoised output min and max value range after clip, by default (-1, 1).
             callback (Callable[[int,int,int,Tensor], None], optional):
                 Callback function during denoising steps.
 
@@ -440,6 +446,9 @@ class GaussianDiffusion(nn.Layer):
 
         # denoising loop
         denoised_output = noisy_input
+        if clip_noise:
+            n_min, n_max = clip_noise_range
+            denoised_output = paddle.clip(denoised_output, n_min, n_max)
         num_warmup_steps = len(
             timesteps) - num_inference_steps * scheduler.order
         for i, t in enumerate(timesteps):
@@ -451,6 +460,8 @@ class GaussianDiffusion(nn.Layer):
             # compute the previous noisy sample x_t -> x_t-1
             denoised_output = scheduler.step(noise_pred, t,
                                              denoised_output).prev_sample
+            if clip_noise:
+                denoised_output = paddle.clip(denoised_output, n_min, n_max)
 
             # call the callback, if provided
             if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and

third_party/ctc_decoders/setup.py

@@ -129,7 +129,7 @@ decoders_module = [
 
 setup(
     name='paddlespeech_ctcdecoders',
-    version='0.2.0',
+    version='0.2.2',
     description="CTC decoders in paddlespeech",
     author="PaddlePaddle Speech and Language Team",
     author_email="paddlesl@baidu.com",