Add dithering to the Speech2TextFeatureExtractor API.

src/transformers/audio_utils.py

@@ -390,6 +390,7 @@ 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 +461,10 @@ 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`, *optional*, defaults to 0.0):
+            Adds dithering. In other words, adds a small Gaussian noise to each frame.
+            E.g. use 4.0 to add dithering with a normal distribution centered
+            around 0.0 with standard deviation 4.0, 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 +545,9 @@ 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 * np.random.randn(*buffer[:frame_length].shape)
+
         if remove_dc_offset:
             buffer[:frame_length] = buffer[:frame_length] - buffer[:frame_length].mean()
 
@@ -591,6 +599,7 @@ 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 +662,10 @@ 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`):
+            Adds dithering. In other words, adds a small Gaussian noise to each frame.
+            E.g. use 4.0 to add dithering with a normal distribution centered
+            around 0.0 with standard deviation 4.0, 0.0 means no dithering.
         preemphasis (`float`, *optional*):
             Applies a pre-emphasis filter to each frame.
         mel_filters (`np.ndarray`, *optional*):
@@ -745,6 +758,9 @@ 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 * np.random.randn(*buffer[:, :frame_length].shape)
+
         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 +52,11 @@ 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):
+            Adds dithering. In other words, adds a small Gaussian noise to each frame.
+            E.g. use 4.0 to add dithering with a normal distribution centered
+            around 0.0 with standard deviation 4.0 (assuming [-32k,+32k] range of kaldi waveform).
+            The value 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 +73,15 @@ 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 +113,12 @@ 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 +130,7 @@ 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",

src/transformers/models/whisper/feature_extraction_whisper.py

@@ -57,6 +57,11 @@ class WhisperFeatureExtractor(SequenceFeatureExtractor):
             Size of the Fourier transform.
         padding_value (`float`, *optional*, defaults to 0.0):
             Padding value used to pad the audio. Should correspond to silences.
+        dither (`float`, *optional*, defaults to 0.0):
+            Adds dithering. In other words, adds a small Gaussian noise to each frame.
+            E.g. use 0.0001 to add dithering with a normal distribution centered
+            around 0.0 with standard deviation 0.0001 (assuming [-1,+1] range of raw_speech).
+            The value 0.0 means no dithering.
     """
 
     model_input_names = ["input_features"]
@@ -69,6 +74,7 @@ def __init__(
         chunk_length=30,
         n_fft=400,
         padding_value=0.0,
+        dither=0.0,
         return_attention_mask=False,  # pad inputs to max length with silence token (zero) and no attention mask
         **kwargs,
     ):
@@ -85,6 +91,7 @@ def __init__(
         self.n_samples = chunk_length * sampling_rate
         self.nb_max_frames = self.n_samples // hop_length
         self.sampling_rate = sampling_rate
+        self.dither = dither
         self.mel_filters = mel_filter_bank(
             num_frequency_bins=1 + n_fft // 2,
             num_mel_filters=feature_size,
@@ -114,6 +121,7 @@ def _np_extract_fbank_features(self, waveform_batch: np.array, device: str) -> n
                 frame_length=self.n_fft,
                 hop_length=self.hop_length,
                 power=2.0,
+                dither=self.dither,
                 mel_filters=self.mel_filters,
                 log_mel="log10",
             )
@@ -135,6 +143,13 @@ def _torch_extract_fbank_features(self, waveform: np.array, device: str = "cpu")
         if device != "cpu":
             waveform = waveform.to(device)
             window = window.to(device)
+
+        # Note: it would be better to dither the chunked waveform,
+        # so overlapping signal does not get the same dithering.
+        # But, chunking is happening inside pytorch, so it is here.
+        if self.dither != 0.0:
+            waveform += self.dither * torch.randn(waveform.shape, dtype=waveform.dtype, device=waveform.device)
+
         stft = torch.stft(waveform, self.n_fft, self.hop_length, window=window, return_complex=True)
         magnitudes = stft[..., :-1].abs() ** 2
 

tests/models/clap/test_feature_extraction_clap.py

@@ -53,7 +53,7 @@ def floats_list(shape, scale=1.0, rng=None, name=None):
 @require_torch
 @require_torchaudio
 # Copied from tests.models.whisper.test_feature_extraction_whisper.WhisperFeatureExtractionTester with Whisper->Clap
-class ClapFeatureExtractionTester(unittest.TestCase):
+class ClapFeatureExtractionTester:
     def __init__(
         self,
         parent,

tests/models/speech_to_text/test_feature_extraction_speech_to_text.py

@@ -48,7 +48,7 @@ def floats_list(shape, scale=1.0, rng=None, name=None):
 
 @require_torch
 @require_torchaudio
-class Speech2TextFeatureExtractionTester(unittest.TestCase):
+class Speech2TextFeatureExtractionTester:
     def __init__(
         self,
         parent,
@@ -144,6 +144,38 @@ def test_call(self):
         for enc_seq_1, enc_seq_2 in zip(encoded_sequences_1, encoded_sequences_2):
             self.assertTrue(np.allclose(enc_seq_1, enc_seq_2, atol=1e-3))
 
+    def test_dither(self):
+        # Tests that features with and without little dithering are similar, but not the same
+        dict_no_dither = self.feat_extract_tester.prepare_feat_extract_dict()
+        dict_no_dither["dither"] = 0.0
+
+        dict_dither = self.feat_extract_tester.prepare_feat_extract_dict()
+        dict_dither["dither"] = 1.0
+
+        feature_extractor_no_dither = self.feature_extraction_class(**dict_no_dither)
+        feature_extractor_dither = self.feature_extraction_class(**dict_dither)
+
+        # create three inputs of length 800, 1000, and 1200
+        speech_inputs = [floats_list((1, x))[0] for x in range(800, 1400, 200)]
+        np_speech_inputs = [np.asarray(speech_input) for speech_input in speech_inputs]
+
+        # compute features
+        input_features_no_dither = feature_extractor_no_dither(
+            np_speech_inputs, padding=True, return_tensors="np"
+        ).input_features
+        input_features_dither = feature_extractor_dither(
+            np_speech_inputs, padding=True, return_tensors="np"
+        ).input_features
+
+        # test there is a difference between features (there's added noise to input signal)
+        diff = input_features_dither - input_features_no_dither
+
+        # features are not identical
+        self.assertTrue(np.abs(diff).mean() > 1e-5)
+        # features are not too different
+        self.assertTrue(np.abs(diff).mean() <= 1e-3)
+        self.assertTrue(np.abs(diff).max() <= 1e-2)
+
     def test_cepstral_mean_and_variance_normalization(self):
         feature_extractor = self.feature_extraction_class(**self.feat_extract_tester.prepare_feat_extract_dict())
         speech_inputs = [floats_list((1, x))[0] for x in range(800, 1400, 200)]

tests/models/whisper/test_feature_extraction_whisper.py

@@ -50,7 +50,7 @@ def floats_list(shape, scale=1.0, rng=None, name=None):
     return values
 
 
-class WhisperFeatureExtractionTester(unittest.TestCase):
+class WhisperFeatureExtractionTester:
     def __init__(
         self,
         parent,
@@ -200,6 +200,38 @@ def test_call(self):
         for enc_seq_1, enc_seq_2 in zip(encoded_sequences_1, encoded_sequences_2):
             self.assertTrue(np.allclose(enc_seq_1, enc_seq_2, atol=1e-3))
 
+    def test_dither(self):
+        # Tests that features with and without little dithering are similar, but not the same
+        dict_no_dither = self.feat_extract_tester.prepare_feat_extract_dict()
+        dict_no_dither["dither"] = 0.0
+
+        dict_dither = self.feat_extract_tester.prepare_feat_extract_dict()
+        dict_dither["dither"] = 0.00003  # approx. 1/32k
+
+        feature_extractor_no_dither = self.feature_extraction_class(**dict_no_dither)
+        feature_extractor_dither = self.feature_extraction_class(**dict_dither)
+
+        # create three inputs of length 800, 1000, and 1200
+        speech_inputs = [floats_list((1, x))[0] for x in range(800, 1400, 200)]
+        np_speech_inputs = [np.asarray(speech_input) for speech_input in speech_inputs]
+
+        # compute features
+        input_features_no_dither = feature_extractor_no_dither(
+            np_speech_inputs, padding=True, return_tensors="np"
+        ).input_features
+        input_features_dither = feature_extractor_dither(
+            np_speech_inputs, padding=True, return_tensors="np"
+        ).input_features
+
+        # test there is a difference between features (there's added noise to input signal)
+        diff = input_features_dither - input_features_no_dither
+
+        # features are not identical
+        self.assertTrue(np.abs(diff).mean() > 1e-6)
+        # features are not too different
+        self.assertTrue(np.abs(diff).mean() <= 1e-4)
+        self.assertTrue(np.abs(diff).max() <= 1e-3)
+
     @require_torch
     def test_double_precision_pad(self):
         import torch