Upload feature extractor

preprocessor_config.json

@@ -3,15 +3,15 @@
     "AutoFeatureExtractor": "processing_bestrq_conformer.ModifiedWhisperFeatureExtractor",
     "AutoProcessor": "processing_bestrq_conformer.ModifiedWhisperFeatureExtractor"
   },
-  "chunk_length": 30,
+  "chunk_length": 120,
   "feature_extractor_type": "ModifiedWhisperFeatureExtractor",
   "feature_size": 80,
   "hop_length": 160,
   "n_fft": 400,
-  "n_samples": 480000,
-  "nb_max_frames": 3000,
+  "n_samples": 1920000,
+  "nb_max_frames": 12000,
   "padding_side": "right",
   "padding_value": 0.0,
-  "return_attention_mask": false,
+  "return_attention_mask": true,
   "sampling_rate": 16000
 }

processing_bestrq_conformer.py

@@ -22,7 +22,6 @@ from shutil import copyfile
 from typing import Dict, List, Optional, Tuple, Union
 
 import numpy as np
-import sentencepiece as spm
 
 from transformers import is_torch_available, AutoFeatureExtractor, AutoTokenizer
 from transformers.audio_utils import mel_filter_bank, spectrogram, window_function
@@ -79,10 +78,10 @@ class ModifiedWhisperFeatureExtractor(SequenceFeatureExtractor):
         feature_size=80,
         sampling_rate=16000,
         hop_length=160,
-        chunk_length=30,
+        chunk_length=120,
         n_fft=400,
         padding_value=0.0,
-        return_attention_mask=False,  # pad inputs to max length with silence token (zero) and no attention mask
+        return_attention_mask=True,  # pad inputs to max length with silence token (zero) and no attention mask
         **kwargs,
     ):
         super().__init__(
@@ -190,7 +189,7 @@ class ModifiedWhisperFeatureExtractor(SequenceFeatureExtractor):
         truncation: bool = True,
         pad_to_multiple_of: Optional[int] = None,
         return_tensors: Optional[Union[str, TensorType]] = None,
-        return_attention_mask: Optional[bool] = None,
+        return_attention_mask: Optional[bool] = True,
         padding: Optional[Union[bool, str]] = True,
         max_length: Optional[int] = None,
         sampling_rate: Optional[int] = None,
@@ -364,6 +363,8 @@ class MeralionBestRqConformerTokenizer(PreTrainedTokenizer):
         blank_token="<blk>",
         **kwargs
     ):
+        import sentencepiece as spm
+
         self.vocab_file = vocab_file
         self.sp_model = spm.SentencePieceProcessor()
         self.sp_model.Load(vocab_file)