[audio utils] fix fft_bin_width computation (#36603)

* fix fft_bin_width computation

* update docstring + enforce correct params

* update test with correct value

* udpate test

* update feature extractors for concerned models

* update

* make

* udpate docstring

* udpate docstring

src/transformers/audio_utils.py

@@ -293,7 +293,7 @@ def mel_filter_bank(
 
     Args:
         num_frequency_bins (`int`):
-            Number of frequencies used to compute the spectrogram (should be the same as in `stft`).
+            Number of frequency bins (should be the same as `n_fft // 2 + 1` where `n_fft` is the size of the Fourier Transform used to compute the spectrogram).
         num_mel_filters (`int`):
             Number of mel filters to generate.
         min_frequency (`float`):
@@ -317,6 +317,12 @@ def mel_filter_bank(
     if norm is not None and norm != "slaney":
         raise ValueError('norm must be one of None or "slaney"')
 
+    if num_frequency_bins < 2:
+        raise ValueError(f"Require num_frequency_bins: {num_frequency_bins} >= 2")
+
+    if min_frequency > max_frequency:
+        raise ValueError(f"Require min_frequency: {min_frequency} <= max_frequency: {max_frequency}")
+
     # center points of the triangular mel filters
     mel_min = hertz_to_mel(min_frequency, mel_scale=mel_scale)
     mel_max = hertz_to_mel(max_frequency, mel_scale=mel_scale)
@@ -325,7 +331,7 @@ def mel_filter_bank(
 
     if triangularize_in_mel_space:
         # frequencies of FFT bins in Hz, but filters triangularized in mel space
-        fft_bin_width = sampling_rate / (num_frequency_bins * 2)
+        fft_bin_width = sampling_rate / ((num_frequency_bins - 1) * 2)
         fft_freqs = hertz_to_mel(fft_bin_width * np.arange(num_frequency_bins), mel_scale=mel_scale)
         filter_freqs = mel_freqs
     else:

src/transformers/models/audio_spectrogram_transformer/feature_extraction_audio_spectrogram_transformer.py

@@ -91,7 +91,7 @@ class ASTFeatureExtractor(SequenceFeatureExtractor):
 
         if not is_speech_available():
             mel_filters = mel_filter_bank(
-                num_frequency_bins=256,
+                num_frequency_bins=257,
                 num_mel_filters=self.num_mel_bins,
                 min_frequency=20,
                 max_frequency=sampling_rate // 2,
@@ -101,7 +101,7 @@ class ASTFeatureExtractor(SequenceFeatureExtractor):
                 triangularize_in_mel_space=True,
             )
 
-            self.mel_filters = np.pad(mel_filters, ((0, 1), (0, 0)))
+            self.mel_filters = mel_filters
             self.window = window_function(400, "hann", periodic=False)
 
     def _extract_fbank_features(

src/transformers/models/seamless_m4t/feature_extraction_seamless_m4t.py

@@ -74,7 +74,7 @@ class SeamlessM4TFeatureExtractor(SequenceFeatureExtractor):
         self.stride = stride
 
         mel_filters = mel_filter_bank(
-            num_frequency_bins=256,
+            num_frequency_bins=257,
             num_mel_filters=self.num_mel_bins,
             min_frequency=20,
             max_frequency=sampling_rate // 2,
@@ -84,7 +84,7 @@ class SeamlessM4TFeatureExtractor(SequenceFeatureExtractor):
             triangularize_in_mel_space=True,
         )
 
-        self.mel_filters = np.pad(mel_filters, ((0, 1), (0, 0)))
+        self.mel_filters = mel_filters
         self.window = window_function(400, "povey", periodic=False)
 
         super().__init__(feature_size=feature_size, sampling_rate=sampling_rate, padding_value=padding_value, **kwargs)

src/transformers/models/speech_to_text/feature_extraction_speech_to_text.py

@@ -91,7 +91,7 @@ class Speech2TextFeatureExtractor(SequenceFeatureExtractor):
 
         if not is_speech_available():
             mel_filters = mel_filter_bank(
-                num_frequency_bins=256,
+                num_frequency_bins=257,
                 num_mel_filters=self.num_mel_bins,
                 min_frequency=20,
                 max_frequency=sampling_rate // 2,
@@ -101,7 +101,7 @@ class Speech2TextFeatureExtractor(SequenceFeatureExtractor):
                 triangularize_in_mel_space=True,
             )
 
-            self.mel_filters = np.pad(mel_filters, ((0, 1), (0, 0)))
+            self.mel_filters = mel_filters
             self.window = window_function(400, "povey", periodic=False)
 
     def _extract_fbank_features(