Add dithering to the Speech2TextFeatureExtractor API.

- in kaldi : https://github.com/kaldi-asr/kaldi/blob/4a8b7f673275597fef8a15b160124bd0985b59bd/src/feat/feature-window.cc#L145
- with dithering without a seed, the features become non-deterministic due
  to small Gaussian noise added to the audio (i.e. 2 runs lead to little
  different outputs)

src/transformers/audio_utils.py

@@ -20,6 +20,8 @@
 import warnings
 from typing import List, Optional, Tuple, Union
 
+import torch
+
 import numpy as np
 
 
@@ -390,6 +392,7 @@ def spectrogram(
     center: bool = True,
     pad_mode: str = "reflect",
     onesided: bool = True,
+    dither: float = 0.0,
     preemphasis: Optional[float] = None,
     mel_filters: Optional[np.ndarray] = None,
     mel_floor: float = 1e-10,
@@ -460,6 +463,9 @@ def spectrogram(
         onesided (`bool`, *optional*, defaults to `True`):
             If True, only computes the positive frequencies and returns a spectrogram containing `fft_length // 2 + 1`
             frequency bins. If False, also computes the negative frequencies and returns `fft_length` frequency bins.
+        dither (`float`):
+            Add dithering (add small Gaussian noise to each frame).
+            E.g. use 4 to add dithering, 0.0 means no dithering.
         preemphasis (`float`, *optional*)
             Coefficient for a low-pass filter that applies pre-emphasis before the DFT.
         mel_filters (`np.ndarray` of shape `(num_freq_bins, num_mel_filters)`, *optional*):
@@ -540,6 +546,13 @@ def spectrogram(
     for frame_idx in range(num_frames):
         buffer[:frame_length] = waveform[timestep : timestep + frame_length]
 
+        if dither != 0.0:
+            buffer[:frame_length] += dither * torch.randn(
+                buffer[:frame_length].shape,
+                device=buffer.device,
+                dtype=buffer.dtype,
+            )
+
         if remove_dc_offset:
             buffer[:frame_length] = buffer[:frame_length] - buffer[:frame_length].mean()
 
@@ -591,6 +604,7 @@ def spectrogram_batch(
     center: bool = True,
     pad_mode: str = "reflect",
     onesided: bool = True,
+    dither: float = 0.0,
     preemphasis: Optional[float] = None,
     mel_filters: Optional[np.ndarray] = None,
     mel_floor: float = 1e-10,
@@ -653,6 +667,9 @@ def spectrogram_batch(
             The padding strategy when `center` is `True`.
         onesided (`bool`, *optional*, defaults to `True`):
             If True, returns a one-sided spectrogram for real input signals.
+        dither (`float`):
+            Add dithering (add small Gaussian noise to each frame).
+            E.g. use 4 to add dithering, 0.0 means no dithering.
         preemphasis (`float`, *optional*):
             Applies a pre-emphasis filter to each frame.
         mel_filters (`np.ndarray`, *optional*):
@@ -745,6 +762,13 @@ def spectrogram_batch(
         timestep = frame_idx * hop_length
         buffer[:, :frame_length] = padded_waveform_batch[:, timestep : timestep + frame_length]
 
+        if dither != 0.0:
+            buffer[:, :frame_length] += dither * torch.randn(
+                buffer[:, :frame_length].shape,
+                device=buffer.device,
+                dtype=buffer.dtype,
+            )
+
         if remove_dc_offset:
             buffer[:, :frame_length] -= buffer[:, :frame_length].mean(axis=1, keepdims=True)
 

src/transformers/models/speech_to_text/feature_extraction_speech_to_text.py

@@ -52,6 +52,9 @@ class Speech2TextFeatureExtractor(SequenceFeatureExtractor):
             Number of Mel-frequency bins.
         padding_value (`float`, *optional*, defaults to 0.0):
             The value that is used to fill the padding vectors.
+        dither (`float`, *optional*, defaults to 0.0):
+            Add dithering (add small Gaussian noise to each frame).
+            E.g. use 4 to add dithering, 0.0 means no dithering.
         do_ceptral_normalize (`bool`, *optional*, defaults to `True`):
             Whether or not to apply utterance-level cepstral mean and variance normalization to extracted features.
         normalize_means (`bool`, *optional*, defaults to `True`):
@@ -68,13 +71,15 @@ def __init__(
         sampling_rate=16000,
         num_mel_bins=80,
         padding_value=0.0,
+        dither=0.0,
         do_ceptral_normalize=True,
         normalize_means=True,
         normalize_vars=True,
         **kwargs,
     ):
         super().__init__(feature_size=feature_size, sampling_rate=sampling_rate, padding_value=padding_value, **kwargs)
         self.num_mel_bins = num_mel_bins
+        self.dither = dither
         self.do_ceptral_normalize = do_ceptral_normalize
         self.normalize_means = normalize_means
         self.normalize_vars = normalize_vars
@@ -106,7 +111,12 @@ def _extract_fbank_features(
         waveform = waveform * (2**15)  # Kaldi compliance: 16-bit signed integers
         if is_speech_available():
             waveform = torch.from_numpy(waveform).unsqueeze(0)
-            features = ta_kaldi.fbank(waveform, num_mel_bins=self.num_mel_bins, sample_frequency=self.sampling_rate)
+            features = ta_kaldi.fbank(
+                waveform,
+                dither=self.dither,
+                num_mel_bins=self.num_mel_bins,
+                sample_frequency=self.sampling_rate,
+            )
             features = features.numpy()
         else:
             waveform = np.squeeze(waveform)
@@ -118,6 +128,7 @@ def _extract_fbank_features(
                 fft_length=512,
                 power=2.0,
                 center=False,
+                dither=self.dither,
                 preemphasis=0.97,
                 mel_filters=self.mel_filters,
                 log_mel="log",