def test_save_load_pretrained_additional_features(self): processor = Speech2TextProcessor( tokenizer=self.get_tokenizer(), feature_extractor=self.get_feature_extractor()) processor.save_pretrained(self.tmpdirname) tokenizer_add_kwargs = self.get_tokenizer(bos_token="(BOS)", eos_token="(EOS)") feature_extractor_add_kwargs = self.get_feature_extractor( do_normalize=False, padding_value=1.0) processor = Speech2TextProcessor.from_pretrained(self.tmpdirname, bos_token="(BOS)", eos_token="(EOS)", do_normalize=False, padding_value=1.0) self.assertEqual(processor.tokenizer.get_vocab(), tokenizer_add_kwargs.get_vocab()) self.assertIsInstance(processor.tokenizer, Speech2TextTokenizer) self.assertEqual(processor.feature_extractor.to_json_string(), feature_extractor_add_kwargs.to_json_string()) self.assertIsInstance(processor.feature_extractor, Speech2TextFeatureExtractor)
def test_save_load_pretrained_default(self): tokenizer = self.get_tokenizer() feature_extractor = self.get_feature_extractor() processor = Speech2TextProcessor(tokenizer=tokenizer, feature_extractor=feature_extractor) processor.save_pretrained(self.tmpdirname) processor = Speech2TextProcessor.from_pretrained(self.tmpdirname) self.assertEqual(processor.tokenizer.get_vocab(), tokenizer.get_vocab()) self.assertIsInstance(processor.tokenizer, Speech2TextTokenizer) self.assertEqual(processor.feature_extractor.to_json_string(), feature_extractor.to_json_string()) self.assertIsInstance(processor.feature_extractor, Speech2TextFeatureExtractor)
def test_tokenizer_decode(self): feature_extractor = self.get_feature_extractor() tokenizer = self.get_tokenizer() processor = Speech2TextProcessor(tokenizer=tokenizer, feature_extractor=feature_extractor) predicted_ids = [[1, 4, 5, 8, 1, 0, 8], [3, 4, 3, 1, 1, 8, 9]] decoded_processor = processor.batch_decode(predicted_ids) decoded_tok = tokenizer.batch_decode(predicted_ids) self.assertListEqual(decoded_tok, decoded_processor)
def test_feature_extractor(self): feature_extractor = self.get_feature_extractor() tokenizer = self.get_tokenizer() processor = Speech2TextProcessor(tokenizer=tokenizer, feature_extractor=feature_extractor) raw_speech = floats_list((3, 1000)) input_feat_extract = feature_extractor(raw_speech, return_tensors="np") input_processor = processor(raw_speech, return_tensors="np") for key in input_feat_extract.keys(): self.assertAlmostEqual(input_feat_extract[key].sum(), input_processor[key].sum(), delta=1e-2)
def s2t_predictions(audio_file): device = 'cuda' if torch.cuda.is_available() else 'cpu' audio_array = s2t_audio_to_array(audio_file) model = Speech2TextForConditionalGeneration.from_pretrained( "facebook/s2t-small-librispeech-asr").to(device).eval() processor = Speech2TextProcessor.from_pretrained( "facebook/s2t-small-librispeech-asr", do_upper_case=True) features = processor(audio_array, sampling_rate=16000, return_tensors="pt") input_features = features.input_features.to(device) attention_mask = features.attention_mask.to(device) gen_tokens = model.generate(input_ids=input_features) text = processor.batch_decode(gen_tokens, skip_special_tokens=True) return text
def test_tokenizer(self): feature_extractor = self.get_feature_extractor() tokenizer = self.get_tokenizer() processor = Speech2TextProcessor(tokenizer=tokenizer, feature_extractor=feature_extractor) input_str = "This is a test string" encoded_processor = processor(text=input_str) encoded_tok = tokenizer(input_str) for key in encoded_tok.keys(): self.assertListEqual(encoded_tok[key], encoded_processor[key])
def translate(data, sampling_rate, pretrained_model_name="facebook/s2t-small-librispeech-asr"): model = Speech2TextForConditionalGeneration.from_pretrained( pretrained_model_name) feature_extractor = Speech2TextFeatureExtractor.from_pretrained( pretrained_model_name) tokenizer = Speech2TextTokenizer.from_pretrained(pretrained_model_name) processor = Speech2TextProcessor(feature_extractor=feature_extractor, tokenizer=tokenizer) inputs = processor(data, sampling_rate=sampling_rate, return_tensors="pt") generated_ids = model.generate(input_ids=inputs["input_features"], attention_mask=inputs["attention_mask"]) transcription = processor.batch_decode(generated_ids) return transcription
def default_processor(self): return Speech2TextProcessor.from_pretrained( "facebook/s2t-small-librispeech-asr")
import numpy as np from scipy.io.wavfile import read from transformers import Speech2TextProcessor, Speech2TextForConditionalGeneration, Speech2TextFeatureExtractor, Speech2TextTokenizer if __name__ == "__main__": audio_path = "../data/review#1.wav" samplerate, data = read(audio_path) data = (data - np.mean(data)) / np.std(data) print(samplerate, len(data)) model = Speech2TextForConditionalGeneration.from_pretrained( "facebook/s2t-small-librispeech-asr") feature_extractor = Speech2TextFeatureExtractor.from_pretrained( "facebook/s2t-small-librispeech-asr") tokenizer = Speech2TextTokenizer.from_pretrained( "facebook/s2t-small-librispeech-asr") processor = Speech2TextProcessor(feature_extractor=feature_extractor, tokenizer=tokenizer) inputs = processor(data, sampling_rate=samplerate, return_tensors="pt") generated_ids = model.generate(input_ids=inputs["input_features"], attention_mask=inputs["attention_mask"]) transcription = processor.batch_decode(generated_ids) print(transcription)