replace kaidi_fbank with paddleaudio

examples/aishell/asr1/conf/preprocess.yaml

@@ -3,8 +3,9 @@ process:
   - type: fbank_kaldi
     fs: 16000
     n_mels: 80
-    n_shift: 160
+    n_frame_length: 25
-    win_length: 400
+    n_frame_shift: 10
+    energy_floor: 0.0
     dither: 0.1
   - type: cmvn_json
     cmvn_path: data/mean_std.json
@@ -23,7 +24,3 @@ process:
     n_mask: 2
     inplace: true
     replace_with_zero: false
-
-
-
-

paddlespeech/s2t/transform/spectrogram.py

@@ -14,8 +14,11 @@
 # Modified from espnet(https://github.com/espnet/espnet)
 import librosa
 import numpy as np
+import paddle
 from python_speech_features import logfbank
 
+import paddleaudio.compliance.kaldi as kaldi
+
 
 def stft(x,
          n_fft,
@@ -309,6 +312,48 @@ class IStft():
 
 
 class LogMelSpectrogramKaldi():
+    def __init__(self,
+                 fs=16000,
+                 n_mels=80,
+                 n_frame_length=25,
+                 n_frame_shift=10,
+                 energy_floor=0.0,
+                 dither=0.1):
+        self.fs = fs
+        self.n_mels = n_mels
+        self.n_frame_length = n_frame_length
+        self.n_frame_shift = n_frame_shift
+        self.energy_floor = energy_floor
+        self.dither = dither
+
+    def __repr__(self):
+        return (
+            "{name}(fs={fs}, n_mels={n_mels}, "
+            "n_frame_shift={n_frame_shift}, n_frame_length={n_frame_length}, "
+            "dither={dither}))".format(
+                name=self.__class__.__name__,
+                fs=self.fs,
+                n_mels=self.n_mels,
+                n_frame_shift=self.n_frame_shift,
+                n_frame_length=self.n_frame_length,
+                dither=self.dither, ))
+
+    def __call__(self, x, train):
+        dither = self.dither if train else 0.0
+        waveform = paddle.to_tensor(np.expand_dims(x, 0), dtype=paddle.float32)
+        mat = kaldi.fbank(
+            waveform,
+            n_mels=self.n_mels,
+            frame_length=self.n_frame_length,
+            frame_shift=self.n_frame_shift,
+            dither=dither,
+            energy_floor=self.energy_floor,
+            sr=self.fs)
+        mat = np.squeeze(mat.numpy())
+        return mat
+
+
+class LogMelSpectrogramKaldi_decay():
     def __init__(
             self,
             fs=16000,

paddlespeech/s2t/transform/transformation.py

@@ -31,6 +31,7 @@ import_alias = dict(
     freq_mask="paddlespeech.s2t.transform.spec_augment:FreqMask",
     spec_augment="paddlespeech.s2t.transform.spec_augment:SpecAugment",
     speed_perturbation="paddlespeech.s2t.transform.perturb:SpeedPerturbation",
+    speed_perturbation_sox="paddlespeech.s2t.transform.perturb:SpeedPerturbationSox",
     volume_perturbation="paddlespeech.s2t.transform.perturb:VolumePerturbation",
     noise_injection="paddlespeech.s2t.transform.perturb:NoiseInjection",
     bandpass_perturbation="paddlespeech.s2t.transform.perturb:BandpassPerturbation",