init

transformers/tests/models/speecht5/test_feature_extraction_speecht5.py

@@ -0,0 +1,420 @@
+# Copyright 2021-2023 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Tests for the SpeechT5 feature extractors."""
+
+import itertools
+import random
+import unittest
+
+import numpy as np
+
+from transformers import BatchFeature, SpeechT5FeatureExtractor
+from transformers.testing_utils import require_torch
+from transformers.utils.import_utils import is_torch_available
+
+from ...test_sequence_feature_extraction_common import SequenceFeatureExtractionTestMixin
+
+
+if is_torch_available():
+    import torch
+
+
+global_rng = random.Random()
+
+
+# Copied from tests.models.whisper.test_feature_extraction_whisper.floats_list
+def floats_list(shape, scale=1.0, rng=None, name=None):
+    """Creates a random float32 tensor"""
+    if rng is None:
+        rng = global_rng
+
+    values = []
+    for batch_idx in range(shape[0]):
+        values.append([])
+        for _ in range(shape[1]):
+            values[-1].append(rng.random() * scale)
+
+    return values
+
+
+@require_torch
+class SpeechT5FeatureExtractionTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=7,
+        min_seq_length=400,
+        max_seq_length=2000,
+        feature_size=1,
+        padding_value=0.0,
+        sampling_rate=16000,
+        do_normalize=True,
+        num_mel_bins=80,
+        hop_length=16,
+        win_length=64,
+        win_function="hann_window",
+        fmin=80,
+        fmax=7600,
+        mel_floor=1e-10,
+        return_attention_mask=True,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.min_seq_length = min_seq_length
+        self.max_seq_length = max_seq_length
+        self.seq_length_diff = (self.max_seq_length - self.min_seq_length) // (self.batch_size - 1)
+        self.feature_size = feature_size
+        self.padding_value = padding_value
+        self.sampling_rate = sampling_rate
+        self.do_normalize = do_normalize
+        self.num_mel_bins = num_mel_bins
+        self.hop_length = hop_length
+        self.win_length = win_length
+        self.win_function = win_function
+        self.fmin = fmin
+        self.fmax = fmax
+        self.mel_floor = mel_floor
+        self.return_attention_mask = return_attention_mask
+
+    def prepare_feat_extract_dict(self):
+        return {
+            "feature_size": self.feature_size,
+            "padding_value": self.padding_value,
+            "sampling_rate": self.sampling_rate,
+            "do_normalize": self.do_normalize,
+            "num_mel_bins": self.num_mel_bins,
+            "hop_length": self.hop_length,
+            "win_length": self.win_length,
+            "win_function": self.win_function,
+            "fmin": self.fmin,
+            "fmax": self.fmax,
+            "mel_floor": self.mel_floor,
+            "return_attention_mask": self.return_attention_mask,
+        }
+
+    def prepare_inputs_for_common(self, equal_length=False, numpify=False):
+        def _flatten(list_of_lists):
+            return list(itertools.chain(*list_of_lists))
+
+        if equal_length:
+            speech_inputs = floats_list((self.batch_size, self.max_seq_length))
+        else:
+            # make sure that inputs increase in size
+            speech_inputs = [
+                _flatten(floats_list((x, self.feature_size)))
+                for x in range(self.min_seq_length, self.max_seq_length, self.seq_length_diff)
+            ]
+
+        if numpify:
+            speech_inputs = [np.asarray(x) for x in speech_inputs]
+
+        return speech_inputs
+
+    def prepare_inputs_for_target(self, equal_length=False, numpify=False):
+        if equal_length:
+            speech_inputs = [floats_list((self.max_seq_length, self.num_mel_bins)) for _ in range(self.batch_size)]
+        else:
+            # make sure that inputs increase in size
+            speech_inputs = [
+                floats_list((x, self.num_mel_bins))
+                for x in range(self.min_seq_length, self.max_seq_length, self.seq_length_diff)
+            ]
+
+        if numpify:
+            speech_inputs = [np.asarray(x) for x in speech_inputs]
+
+        return speech_inputs
+
+
+@require_torch
+class SpeechT5FeatureExtractionTest(SequenceFeatureExtractionTestMixin, unittest.TestCase):
+    feature_extraction_class = SpeechT5FeatureExtractor
+
+    def setUp(self):
+        self.feat_extract_tester = SpeechT5FeatureExtractionTester(self)
+
+    def _check_zero_mean_unit_variance(self, input_vector):
+        self.assertTrue(np.all(np.mean(input_vector, axis=0) < 1e-3))
+        self.assertTrue(np.all(np.abs(np.var(input_vector, axis=0) - 1) < 1e-3))
+
+    def test_call(self):
+        # Tests that all call wrap to encode_plus and batch_encode_plus
+        feat_extract = self.feature_extraction_class(**self.feat_extract_tester.prepare_feat_extract_dict())
+        # 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]
+
+        # Test not batched input
+        encoded_sequences_1 = feat_extract(speech_inputs[0], return_tensors="np").input_values
+        encoded_sequences_2 = feat_extract(np_speech_inputs[0], return_tensors="np").input_values
+        self.assertTrue(np.allclose(encoded_sequences_1, encoded_sequences_2, atol=1e-3))
+
+        # Test batched
+        encoded_sequences_1 = feat_extract(speech_inputs, return_tensors="np").input_values
+        encoded_sequences_2 = feat_extract(np_speech_inputs, return_tensors="np").input_values
+        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_zero_mean_unit_variance_normalization_np(self):
+        feat_extract = 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)]
+
+        paddings = ["longest", "max_length", "do_not_pad"]
+        max_lengths = [None, 1600, None]
+        for max_length, padding in zip(max_lengths, paddings):
+            processed = feat_extract(speech_inputs, padding=padding, max_length=max_length, return_tensors="np")
+            input_values = processed.input_values
+
+            self._check_zero_mean_unit_variance(input_values[0][:800])
+            self.assertTrue(input_values[0][800:].sum() < 1e-6)
+            self._check_zero_mean_unit_variance(input_values[1][:1000])
+            self.assertTrue(input_values[0][1000:].sum() < 1e-6)
+            self._check_zero_mean_unit_variance(input_values[2][:1200])
+
+    def test_zero_mean_unit_variance_normalization(self):
+        feat_extract = self.feature_extraction_class(**self.feat_extract_tester.prepare_feat_extract_dict())
+        lengths = range(800, 1400, 200)
+        speech_inputs = [floats_list((1, x))[0] for x in lengths]
+
+        paddings = ["longest", "max_length", "do_not_pad"]
+        max_lengths = [None, 1600, None]
+
+        for max_length, padding in zip(max_lengths, paddings):
+            processed = feat_extract(speech_inputs, max_length=max_length, padding=padding)
+            input_values = processed.input_values
+
+            self._check_zero_mean_unit_variance(input_values[0][:800])
+            self._check_zero_mean_unit_variance(input_values[1][:1000])
+            self._check_zero_mean_unit_variance(input_values[2][:1200])
+
+    def test_zero_mean_unit_variance_normalization_trunc_np_max_length(self):
+        feat_extract = 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)]
+        processed = feat_extract(
+            speech_inputs, truncation=True, max_length=1000, padding="max_length", return_tensors="np"
+        )
+        input_values = processed.input_values
+
+        self._check_zero_mean_unit_variance(input_values[0, :800])
+        self._check_zero_mean_unit_variance(input_values[1])
+        self._check_zero_mean_unit_variance(input_values[2])
+
+    def test_zero_mean_unit_variance_normalization_trunc_np_longest(self):
+        feat_extract = 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)]
+        processed = feat_extract(
+            speech_inputs, truncation=True, max_length=1000, padding="longest", return_tensors="np"
+        )
+        input_values = processed.input_values
+
+        self._check_zero_mean_unit_variance(input_values[0, :800])
+        self._check_zero_mean_unit_variance(input_values[1, :1000])
+        self._check_zero_mean_unit_variance(input_values[2])
+
+        # make sure that if max_length < longest -> then pad to max_length
+        self.assertTrue(input_values.shape == (3, 1000))
+
+        speech_inputs = [floats_list((1, x))[0] for x in range(800, 1400, 200)]
+        processed = feat_extract(
+            speech_inputs, truncation=True, max_length=2000, padding="longest", return_tensors="np"
+        )
+        input_values = processed.input_values
+
+        self._check_zero_mean_unit_variance(input_values[0, :800])
+        self._check_zero_mean_unit_variance(input_values[1, :1000])
+        self._check_zero_mean_unit_variance(input_values[2])
+
+        # make sure that if max_length > longest -> then pad to longest
+        self.assertTrue(input_values.shape == (3, 1200))
+
+    def test_double_precision_pad(self):
+        feature_extractor = self.feature_extraction_class(**self.feat_extract_tester.prepare_feat_extract_dict())
+        np_speech_inputs = np.random.rand(100).astype(np.float64)
+        py_speech_inputs = np_speech_inputs.tolist()
+
+        for inputs in [py_speech_inputs, np_speech_inputs]:
+            np_processed = feature_extractor.pad([{"input_values": inputs}], return_tensors="np")
+            self.assertTrue(np_processed.input_values.dtype == np.float32)
+            pt_processed = feature_extractor.pad([{"input_values": inputs}], return_tensors="pt")
+            self.assertTrue(pt_processed.input_values.dtype == torch.float32)
+
+    def test_call_target(self):
+        # Tests that all call wrap to encode_plus and batch_encode_plus
+        feature_extractor = self.feature_extraction_class(**self.feat_extract_tester.prepare_feat_extract_dict())
+        # 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]
+
+        # Test feature size
+        input_values = feature_extractor(audio_target=np_speech_inputs, padding=True, return_tensors="np").input_values
+        self.assertTrue(input_values.ndim == 3)
+        self.assertTrue(input_values.shape[-1] == feature_extractor.num_mel_bins)
+
+        # Test not batched input
+        encoded_sequences_1 = feature_extractor(speech_inputs[0], return_tensors="np").input_values
+        encoded_sequences_2 = feature_extractor(np_speech_inputs[0], return_tensors="np").input_values
+        self.assertTrue(np.allclose(encoded_sequences_1, encoded_sequences_2, atol=1e-3))
+
+        # Test batched
+        encoded_sequences_1 = feature_extractor(speech_inputs, return_tensors="np").input_values
+        encoded_sequences_2 = feature_extractor(np_speech_inputs, return_tensors="np").input_values
+        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))
+
+        # Test 2-D numpy arrays are batched.
+        speech_inputs = [floats_list((1, x))[0] for x in (800, 800, 800)]
+        np_speech_inputs = np.asarray(speech_inputs)
+        encoded_sequences_1 = feature_extractor(speech_inputs, return_tensors="np").input_values
+        encoded_sequences_2 = feature_extractor(np_speech_inputs, return_tensors="np").input_values
+        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_batch_feature_target(self):
+        speech_inputs = self.feat_extract_tester.prepare_inputs_for_target()
+        feat_extract = self.feature_extraction_class(**self.feat_extract_dict)
+        input_name = feat_extract.model_input_names[0]
+
+        processed_features = BatchFeature({input_name: speech_inputs})
+
+        self.assertTrue(all(len(x) == len(y) for x, y in zip(speech_inputs, processed_features[input_name])))
+
+        speech_inputs = self.feat_extract_tester.prepare_inputs_for_target(equal_length=True)
+        processed_features = BatchFeature({input_name: speech_inputs}, tensor_type="np")
+
+        batch_features_input = processed_features[input_name]
+
+        if len(batch_features_input.shape) < 3:
+            batch_features_input = batch_features_input[:, :, None]
+
+        self.assertTrue(
+            batch_features_input.shape
+            == (self.feat_extract_tester.batch_size, len(speech_inputs[0]), self.feat_extract_tester.num_mel_bins)
+        )
+
+    @require_torch
+    def test_batch_feature_target_pt(self):
+        speech_inputs = self.feat_extract_tester.prepare_inputs_for_target(equal_length=True)
+        feat_extract = self.feature_extraction_class(**self.feat_extract_dict)
+        input_name = feat_extract.model_input_names[0]
+
+        processed_features = BatchFeature({input_name: speech_inputs}, tensor_type="pt")
+
+        batch_features_input = processed_features[input_name]
+
+        if len(batch_features_input.shape) < 3:
+            batch_features_input = batch_features_input[:, :, None]
+
+        self.assertTrue(
+            batch_features_input.shape
+            == (self.feat_extract_tester.batch_size, len(speech_inputs[0]), self.feat_extract_tester.num_mel_bins)
+        )
+
+    @require_torch
+    def test_padding_accepts_tensors_target_pt(self):
+        feat_extract = self.feature_extraction_class(**self.feat_extract_dict)
+        speech_inputs = self.feat_extract_tester.prepare_inputs_for_target()
+        input_name = feat_extract.model_input_names[0]
+
+        processed_features = BatchFeature({input_name: speech_inputs})
+
+        feat_extract.feature_size = feat_extract.num_mel_bins  # hack!
+
+        input_np = feat_extract.pad(processed_features, padding="longest", return_tensors="np")[input_name]
+        input_pt = feat_extract.pad(processed_features, padding="longest", return_tensors="pt")[input_name]
+
+        self.assertTrue(abs(input_np.astype(np.float32).sum() - input_pt.numpy().astype(np.float32).sum()) < 1e-2)
+
+    def test_attention_mask_target(self):
+        feat_dict = self.feat_extract_dict
+        feat_dict["return_attention_mask"] = True
+        feat_extract = self.feature_extraction_class(**feat_dict)
+        speech_inputs = self.feat_extract_tester.prepare_inputs_for_target()
+        input_lengths = [len(x) for x in speech_inputs]
+        input_name = feat_extract.model_input_names[0]
+
+        processed = BatchFeature({input_name: speech_inputs})
+
+        feat_extract.feature_size = feat_extract.num_mel_bins  # hack!
+
+        processed = feat_extract.pad(processed, padding="longest", return_tensors="np")
+        self.assertIn("attention_mask", processed)
+        self.assertListEqual(list(processed.attention_mask.shape), list(processed[input_name].shape[:2]))
+        self.assertListEqual(processed.attention_mask.sum(-1).tolist(), input_lengths)
+
+    def test_attention_mask_with_truncation_target(self):
+        feat_dict = self.feat_extract_dict
+        feat_dict["return_attention_mask"] = True
+        feat_extract = self.feature_extraction_class(**feat_dict)
+        speech_inputs = self.feat_extract_tester.prepare_inputs_for_target()
+        input_lengths = [len(x) for x in speech_inputs]
+        input_name = feat_extract.model_input_names[0]
+
+        processed = BatchFeature({input_name: speech_inputs})
+        max_length = min(input_lengths)
+
+        feat_extract.feature_size = feat_extract.num_mel_bins  # hack!
+
+        processed_pad = feat_extract.pad(
+            processed, padding="max_length", max_length=max_length, truncation=True, return_tensors="np"
+        )
+        self.assertIn("attention_mask", processed_pad)
+        self.assertListEqual(
+            list(processed_pad.attention_mask.shape), [processed_pad[input_name].shape[0], max_length]
+        )
+        self.assertListEqual(
+            processed_pad.attention_mask[:, :max_length].sum(-1).tolist(), [max_length for x in speech_inputs]
+        )
+
+    def _load_datasamples(self, num_samples):
+        from datasets import load_dataset
+
+        ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+        # automatic decoding with librispeech
+        speech_samples = ds.sort("id")[:num_samples]["audio"]
+
+        return [x["array"] for x in speech_samples]
+
+    def test_integration(self):
+        # fmt: off
+        EXPECTED_INPUT_VALUES = torch.tensor(
+            [2.3804e-03, 2.0752e-03, 1.9836e-03, 2.1057e-03, 1.6174e-03,
+             3.0518e-04, 9.1553e-05, 3.3569e-04, 9.7656e-04, 1.8311e-03,
+             2.0142e-03, 2.1057e-03, 1.7395e-03, 4.5776e-04, -3.9673e-04,
+             4.5776e-04, 1.0071e-03, 9.1553e-05, 4.8828e-04, 1.1597e-03,
+             7.3242e-04, 9.4604e-04, 1.8005e-03, 1.8311e-03, 8.8501e-04,
+             4.2725e-04, 4.8828e-04, 7.3242e-04, 1.0986e-03, 2.1057e-03]
+        )
+        # fmt: on
+
+        input_speech = self._load_datasamples(1)
+        feature_extractor = SpeechT5FeatureExtractor()
+        input_values = feature_extractor(input_speech, return_tensors="pt").input_values
+        self.assertEqual(input_values.shape, (1, 93680))
+        torch.testing.assert_close(input_values[0, :30], EXPECTED_INPUT_VALUES, rtol=1e-6, atol=1e-6)
+
+    def test_integration_target(self):
+        # fmt: off
+        EXPECTED_INPUT_VALUES = torch.tensor(
+            [-2.6870, -3.0104, -3.1356, -3.5352, -3.0044, -3.0353, -3.4719, -3.6777,
+             -3.1520, -2.9435, -2.6553, -2.8795, -2.9944, -2.5921, -3.0279, -3.0386,
+             -3.0864, -3.1291, -3.2353, -2.7444, -2.6831, -2.7287, -3.1761, -3.1571,
+             -3.2726, -3.0582, -3.1007, -3.4533, -3.4695, -3.0998]
+        )
+        # fmt: on
+
+        input_speech = self._load_datasamples(1)
+        feature_extractor = SpeechT5FeatureExtractor()
+        input_values = feature_extractor(audio_target=input_speech, return_tensors="pt").input_values
+        self.assertEqual(input_values.shape, (1, 366, 80))
+        torch.testing.assert_close(input_values[0, 0, :30], EXPECTED_INPUT_VALUES, rtol=1e-4, atol=1e-4)

transformers/tests/models/speecht5/test_processing_speecht5.py

@@ -0,0 +1,181 @@
+# Copyright 2022 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Tests for the SpeechT5 processors."""
+
+import json
+import os
+import shutil
+import tempfile
+import unittest
+
+from transformers import is_speech_available, is_torch_available
+from transformers.models.speecht5 import SpeechT5Tokenizer
+from transformers.testing_utils import get_tests_dir, require_speech, require_torch
+from transformers.utils import FEATURE_EXTRACTOR_NAME
+
+
+if is_speech_available() and is_torch_available():
+    from transformers import SpeechT5FeatureExtractor, SpeechT5Processor
+
+    from .test_feature_extraction_speecht5 import floats_list
+
+
+SAMPLE_VOCAB = get_tests_dir("fixtures/test_sentencepiece_bpe_char.model")
+
+
+@require_torch
+@require_speech
+class SpeechT5ProcessorTest(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.tmpdirname = tempfile.mkdtemp()
+
+        tokenizer = SpeechT5Tokenizer(SAMPLE_VOCAB)
+        tokenizer.save_pretrained(cls.tmpdirname)
+
+        feature_extractor_map = {
+            "feature_size": 1,
+            "padding_value": 0.0,
+            "sampling_rate": 16000,
+            "do_normalize": False,
+            "num_mel_bins": 80,
+            "hop_length": 16,
+            "win_length": 64,
+            "win_function": "hann_window",
+            "fmin": 80,
+            "fmax": 7600,
+            "mel_floor": 1e-10,
+            "reduction_factor": 2,
+            "return_attention_mask": True,
+        }
+
+        cls.feature_extraction_file = os.path.join(cls.tmpdirname, FEATURE_EXTRACTOR_NAME)
+        with open(cls.feature_extraction_file, "w", encoding="utf-8") as fp:
+            fp.write(json.dumps(feature_extractor_map) + "\n")
+
+    def get_tokenizer(self, **kwargs):
+        return SpeechT5Tokenizer.from_pretrained(self.tmpdirname, **kwargs)
+
+    def get_feature_extractor(self, **kwargs):
+        return SpeechT5FeatureExtractor.from_pretrained(self.tmpdirname, **kwargs)
+
+    @classmethod
+    def tearDownClass(cls):
+        shutil.rmtree(cls.tmpdirname, ignore_errors=True)
+
+    def test_save_load_pretrained_default(self):
+        tokenizer = self.get_tokenizer()
+        feature_extractor = self.get_feature_extractor()
+
+        processor = SpeechT5Processor(tokenizer=tokenizer, feature_extractor=feature_extractor)
+
+        processor.save_pretrained(self.tmpdirname)
+        processor = SpeechT5Processor.from_pretrained(self.tmpdirname)
+
+        self.assertEqual(processor.tokenizer.get_vocab(), tokenizer.get_vocab())
+        self.assertIsInstance(processor.tokenizer, SpeechT5Tokenizer)
+
+        self.assertEqual(processor.feature_extractor.to_json_string(), feature_extractor.to_json_string())
+        self.assertIsInstance(processor.feature_extractor, SpeechT5FeatureExtractor)
+
+    def test_save_load_pretrained_additional_features(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            processor = SpeechT5Processor(
+                tokenizer=self.get_tokenizer(), feature_extractor=self.get_feature_extractor()
+            )
+            processor.save_pretrained(tmpdir)
+
+            tokenizer_add_kwargs = SpeechT5Tokenizer.from_pretrained(tmpdir, bos_token="(BOS)", eos_token="(EOS)")
+            feature_extractor_add_kwargs = SpeechT5FeatureExtractor.from_pretrained(
+                tmpdir, do_normalize=False, padding_value=1.0
+            )
+
+            processor = SpeechT5Processor.from_pretrained(
+                tmpdir, 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, SpeechT5Tokenizer)
+
+        self.assertEqual(processor.feature_extractor.to_json_string(), feature_extractor_add_kwargs.to_json_string())
+        self.assertIsInstance(processor.feature_extractor, SpeechT5FeatureExtractor)
+
+    def test_feature_extractor(self):
+        feature_extractor = self.get_feature_extractor()
+        tokenizer = self.get_tokenizer()
+
+        processor = SpeechT5Processor(tokenizer=tokenizer, feature_extractor=feature_extractor)
+
+        raw_speech = floats_list((3, 1000))
+
+        input_feat_extract = feature_extractor(audio=raw_speech, return_tensors="np")
+        input_processor = processor(audio=raw_speech, return_tensors="np")
+
+        for key in input_feat_extract:
+            self.assertAlmostEqual(input_feat_extract[key].sum(), input_processor[key].sum(), delta=1e-2)
+
+    def test_feature_extractor_target(self):
+        feature_extractor = self.get_feature_extractor()
+        tokenizer = self.get_tokenizer()
+
+        processor = SpeechT5Processor(tokenizer=tokenizer, feature_extractor=feature_extractor)
+
+        raw_speech = floats_list((3, 1000))
+
+        input_feat_extract = feature_extractor(audio_target=raw_speech, return_tensors="np")
+        input_processor = processor(audio_target=raw_speech, return_tensors="np")
+
+        for key in input_feat_extract:
+            self.assertAlmostEqual(input_feat_extract[key].sum(), input_processor[key].sum(), delta=1e-2)
+
+    def test_tokenizer(self):
+        feature_extractor = self.get_feature_extractor()
+        tokenizer = self.get_tokenizer()
+
+        processor = SpeechT5Processor(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:
+            self.assertListEqual(encoded_tok[key], encoded_processor[key])
+
+    def test_tokenizer_target(self):
+        feature_extractor = self.get_feature_extractor()
+        tokenizer = self.get_tokenizer()
+
+        processor = SpeechT5Processor(tokenizer=tokenizer, feature_extractor=feature_extractor)
+
+        input_str = "This is a test string"
+
+        encoded_processor = processor(text_target=input_str)
+        encoded_tok = tokenizer(input_str)
+
+        for key in encoded_tok:
+            self.assertListEqual(encoded_tok[key], encoded_processor[key])
+
+    def test_tokenizer_decode(self):
+        feature_extractor = self.get_feature_extractor()
+        tokenizer = self.get_tokenizer()
+
+        processor = SpeechT5Processor(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)

transformers/tests/models/speecht5/test_tokenization_speecht5.py

@@ -0,0 +1,221 @@
+# Copyright 2022 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Tests for the SpeechT5 tokenizers."""
+
+import unittest
+
+from transformers import SPIECE_UNDERLINE
+from transformers.models.speecht5 import SpeechT5Tokenizer
+from transformers.testing_utils import get_tests_dir, require_sentencepiece, require_tokenizers, slow
+from transformers.tokenization_utils import AddedToken
+
+from ...test_tokenization_common import TokenizerTesterMixin
+
+
+SAMPLE_VOCAB = get_tests_dir("fixtures/test_sentencepiece_bpe_char.model")
+
+
+@require_sentencepiece
+@require_tokenizers
+class SpeechT5TokenizerTest(TokenizerTesterMixin, unittest.TestCase):
+    from_pretrained_id = "microsoft/speecht5_asr"
+    tokenizer_class = SpeechT5Tokenizer
+    test_rust_tokenizer = False
+    test_sentencepiece = True
+
+    @classmethod
+    def setUpClass(cls):
+        super().setUpClass()
+
+        # We have a SentencePiece fixture for testing
+        tokenizer = SpeechT5Tokenizer(SAMPLE_VOCAB)
+
+        mask_token = AddedToken("<mask>", lstrip=True, rstrip=False)
+        tokenizer.mask_token = mask_token
+        tokenizer.add_special_tokens({"mask_token": mask_token})
+        tokenizer.add_tokens(["<ctc_blank>"])
+
+        tokenizer.save_pretrained(cls.tmpdirname)
+
+    def get_input_output_texts(self, tokenizer):
+        input_text = "this is a test"
+        output_text = "this is a test"
+        return input_text, output_text
+
+    def get_numeric_input_output_texts(self):
+        input_text = "I have $123.45 and owe €59.78. My balance is -₴876.90 and have 73% stocks in my company which equals to ₦72649201"
+        output_text = "I have one hundred and twenty three point four five dollars and owe fifty nine point seven eight euros. My balance is minus eight hundred and seventy six point nine zero ukrainian hryvnia and have seventy three percent stocks in my company which equals to seventy two million six hundred and forty nine thousand two hundred and one nigerian naira"
+        return input_text, output_text
+
+    def get_clean_sequence(self, tokenizer, with_prefix_space=False, max_length=20, min_length=5):
+        input_text, output_text = self.get_input_output_texts(tokenizer)
+        ids = tokenizer.encode(output_text, add_special_tokens=False)
+        text = tokenizer.decode(ids, clean_up_tokenization_spaces=False)
+        return text, ids
+
+    def test_tokenizer_normalization(self):
+        tokenizer = self.get_tokenizer(normalize=True)
+        input_text, expected_text = self.get_numeric_input_output_texts()
+        input_ids = tokenizer.encode(input_text)
+        output_text = tokenizer.decode(input_ids, skip_special_tokens=True)
+        self.assertEqual(output_text, expected_text)
+
+    def test_convert_token_and_id(self):
+        """Test ``_convert_token_to_id`` and ``_convert_id_to_token``."""
+        token = "<pad>"
+        token_id = 1
+
+        self.assertEqual(self.get_tokenizer()._convert_token_to_id(token), token_id)
+        self.assertEqual(self.get_tokenizer()._convert_id_to_token(token_id), token)
+
+    def test_get_vocab(self):
+        vocab_keys = list(self.get_tokenizer().get_vocab().keys())
+
+        self.assertEqual(vocab_keys[0], "<s>")
+        self.assertEqual(vocab_keys[1], "<pad>")
+        self.assertEqual(vocab_keys[-4], "œ")
+        self.assertEqual(vocab_keys[-2], "<mask>")
+        self.assertEqual(vocab_keys[-1], "<ctc_blank>")
+        self.assertEqual(len(vocab_keys), 81)
+
+    def test_vocab_size(self):
+        self.assertEqual(self.get_tokenizer().vocab_size, 79)
+
+    def test_add_tokens_tokenizer(self):
+        tokenizers = self.get_tokenizers(do_lower_case=False)
+        for tokenizer in tokenizers:
+            with self.subTest(f"{tokenizer.__class__.__name__}"):
+                vocab_size = tokenizer.vocab_size
+                all_size = len(tokenizer)
+
+                self.assertNotEqual(vocab_size, 0)
+
+                # We usually have added tokens from the start in tests because our vocab fixtures are
+                # smaller than the original vocabs - let's not assert this
+                # self.assertEqual(vocab_size, all_size)
+
+                new_toks = ["aaaaa bbbbbb", "cccccccccdddddddd"]
+                added_toks = tokenizer.add_tokens(new_toks)
+                vocab_size_2 = tokenizer.vocab_size
+                all_size_2 = len(tokenizer)
+
+                self.assertNotEqual(vocab_size_2, 0)
+                self.assertEqual(vocab_size, vocab_size_2)
+                self.assertEqual(added_toks, len(new_toks))
+                self.assertEqual(all_size_2, all_size + len(new_toks))
+
+                tokens = tokenizer.encode("aaaaa bbbbbb low cccccccccdddddddd l", add_special_tokens=False)
+
+                self.assertGreaterEqual(len(tokens), 4)
+                self.assertGreater(tokens[0], tokenizer.vocab_size - 1)
+                self.assertGreater(tokens[-3], tokenizer.vocab_size - 1)
+
+                new_toks_2 = {"eos_token": ">>>>|||<||<<|<<", "pad_token": "<<<<<|||>|>>>>|>"}
+                added_toks_2 = tokenizer.add_special_tokens(new_toks_2)
+                vocab_size_3 = tokenizer.vocab_size
+                all_size_3 = len(tokenizer)
+
+                self.assertNotEqual(vocab_size_3, 0)
+                self.assertEqual(vocab_size, vocab_size_3)
+                self.assertEqual(added_toks_2, len(new_toks_2))
+                self.assertEqual(all_size_3, all_size_2 + len(new_toks_2))
+
+                tokens = tokenizer.encode(
+                    ">>>>|||<||<<|<< aaaaabbbbbb low cccccccccdddddddd <<<<<|||>|>>>>|> l", add_special_tokens=False
+                )
+
+                self.assertGreaterEqual(len(tokens), 6)
+                self.assertGreater(tokens[0], tokenizer.vocab_size - 1)
+                self.assertGreater(tokens[0], tokens[1])
+                self.assertGreater(tokens[-3], tokenizer.vocab_size - 1)
+                self.assertGreater(tokens[-3], tokens[-4])
+                self.assertEqual(tokens[0], tokenizer.eos_token_id)
+                self.assertEqual(tokens[-3], tokenizer.pad_token_id)
+
+    @unittest.skip
+    def test_pickle_subword_regularization_tokenizer(self):
+        pass
+
+    @unittest.skip
+    def test_subword_regularization_tokenizer(self):
+        pass
+
+    def test_full_tokenizer(self):
+        tokenizer = self.get_tokenizer(normalize=True)
+
+        tokens = tokenizer.tokenize("This is a test")
+        self.assertListEqual(tokens, [SPIECE_UNDERLINE, 'T', 'h', 'i', 's', SPIECE_UNDERLINE, 'i', 's', SPIECE_UNDERLINE, 'a', SPIECE_UNDERLINE, 't', 'e', 's', 't'])  # fmt: skip
+
+        self.assertListEqual(
+            tokenizer.convert_tokens_to_ids(tokens),
+            [4, 32, 11, 10, 12, 4, 10, 12, 4, 7, 4, 6, 5, 12, 6],
+        )
+
+        tokens = tokenizer.tokenize("I was born in 92000, and this is falsé.")
+        self.assertListEqual(tokens,[SPIECE_UNDERLINE, 'I', SPIECE_UNDERLINE, 'w', 'a', 's', SPIECE_UNDERLINE, 'b', 'o', 'r', 'n', SPIECE_UNDERLINE, 'i', 'n', SPIECE_UNDERLINE, 'n', 'i', 'n', 'e', 't', 'y', SPIECE_UNDERLINE, 't', 'w', 'o', SPIECE_UNDERLINE, 't', 'h', 'o', 'u', 's', 'a', 'n', 'd', ',', SPIECE_UNDERLINE, 'a', 'n', 'd', SPIECE_UNDERLINE, 't', 'h', 'i', 's', SPIECE_UNDERLINE, 'i', 's', SPIECE_UNDERLINE, 'f', 'a', 'l', 's', 'é', '.'])  # fmt: skip
+
+        ids = tokenizer.convert_tokens_to_ids(tokens)
+        self.assertListEqual(ids, [4, 30, 4, 20, 7, 12, 4, 25, 8, 13, 9, 4, 10, 9, 4, 9, 10, 9, 5, 6, 22, 4, 6, 20, 8, 4, 6, 11, 8, 16, 12, 7, 9, 14, 23, 4, 7, 9, 14, 4, 6, 11, 10, 12, 4, 10, 12, 4, 19, 7, 15, 12, 73, 26])  # fmt: skip
+
+        back_tokens = tokenizer.convert_ids_to_tokens(ids)
+        self.assertListEqual(back_tokens,[SPIECE_UNDERLINE, 'I', SPIECE_UNDERLINE, 'w', 'a', 's', SPIECE_UNDERLINE, 'b', 'o', 'r', 'n', SPIECE_UNDERLINE, 'i', 'n', SPIECE_UNDERLINE, 'n', 'i', 'n', 'e', 't', 'y', SPIECE_UNDERLINE, 't', 'w', 'o', SPIECE_UNDERLINE, 't', 'h', 'o', 'u', 's', 'a', 'n', 'd', ',', SPIECE_UNDERLINE, 'a', 'n', 'd', SPIECE_UNDERLINE, 't', 'h', 'i', 's', SPIECE_UNDERLINE, 'i', 's', SPIECE_UNDERLINE, 'f', 'a', 'l', 's', 'é', '.'])  # fmt: skip
+
+    @slow
+    def test_tokenizer_integration(self):
+        # Use custom sequence because this tokenizer does not handle numbers.
+        sequences = [
+            "Transformers (formerly known as pytorch-transformers and pytorch-pretrained-bert) provides "
+            "general-purpose architectures (BERT, GPT, RoBERTa, XLM, DistilBert, XLNet...) for Natural "
+            "Language Understanding (NLU) and Natural Language Generation (NLG) with over thirty-two pretrained "
+            "models in one hundred plus languages and deep interoperability between Jax, PyTorch and TensorFlow.",
+            "BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly "
+            "conditioning on both left and right context in all layers.",
+            "The quick brown fox jumps over the lazy dog.",
+        ]
+
+        # fmt: off
+        expected_encoding = {
+            'input_ids': [
+                [4, 32, 13, 7, 9, 12, 19, 8, 13, 18, 5, 13, 12, 4, 64, 19, 8, 13, 18, 5, 13, 15, 22, 4, 28, 9, 8, 20, 9, 4, 7, 12, 4, 24, 22, 6, 8, 13, 17, 11, 39, 6, 13, 7, 9, 12, 19, 8, 13, 18, 5, 13, 12, 4, 7, 9, 14, 4, 24, 22, 6, 8, 13, 17, 11, 39, 24, 13, 5, 6, 13, 7, 10, 9, 5, 14, 39, 25, 5, 13, 6, 63, 4, 24, 13, 8, 27, 10, 14, 5, 12, 4, 21, 5, 9, 5, 13, 7, 15, 39, 24, 16, 13, 24, 8, 12, 5, 4, 7, 13, 17, 11, 10, 6, 5, 17, 6, 16, 13, 5, 12, 4, 64, 40, 47, 54, 32, 23, 4, 53, 49, 32, 23, 4, 54, 8, 40, 47, 54, 32, 7, 23, 4, 69, 52, 43, 23, 4, 51, 10, 12, 6, 10, 15, 40, 5, 13, 6, 23, 4, 69, 52, 48, 5, 6, 26, 26, 26, 63, 4, 19, 8, 13, 4, 48, 7, 6, 16, 13, 7, 15, 4, 52, 7, 9, 21, 16, 7, 21, 5, 4, 61, 9, 14, 5, 13, 12, 6, 7, 9, 14, 10, 9, 21, 4, 64, 48, 52, 61, 63, 4, 7, 9, 14, 4, 48, 7, 6, 16, 13, 7, 15, 4, 52, 7, 9, 21, 16, 7, 21, 5, 4, 53, 5, 9, 5, 13, 7, 6, 10, 8, 9, 4, 64, 48, 52, 53, 63, 4, 20, 10, 6, 11, 4, 8, 27, 5, 13, 4, 6, 11, 10, 13, 6, 22, 39, 6, 20, 8, 4, 24, 13, 5, 6, 13, 7, 10, 9, 5, 14, 4, 18, 8, 14, 5, 15, 12, 4, 10, 9, 4, 8, 9, 5, 4, 11, 16, 9, 14, 13, 5, 14, 4, 24, 15, 16, 12, 4, 15, 7, 9, 21, 16, 7, 21, 5, 12, 4, 7, 9, 14, 4, 14, 5, 5, 24, 4, 10, 9, 6, 5, 13, 8, 24, 5, 13, 7, 25, 10, 15, 10, 6, 22, 4, 25, 5, 6, 20, 5, 5, 9, 4, 58, 7, 37, 23, 4, 49, 22, 32, 8, 13, 17, 11, 4, 7, 9, 14, 4, 32, 5, 9, 12, 8, 13, 55, 15, 8, 20, 26, 2],
+                [4, 40, 47, 54, 32, 4, 10, 12, 4, 14, 5, 12, 10, 21, 9, 5, 14, 4, 6, 8, 4, 24, 13, 5, 39, 6, 13, 7, 10, 9, 4, 14, 5, 5, 24, 4, 25, 10, 14, 10, 13, 5, 17, 6, 10, 8, 9, 7, 15, 4, 13, 5, 24, 13, 5, 12, 5, 9, 6, 7, 6, 10, 8, 9, 12, 4, 19, 13, 8, 18, 4, 16, 9, 15, 7, 25, 5, 15, 5, 14, 4, 6, 5, 37, 6, 4, 25, 22, 4, 46, 8, 10, 9, 6, 15, 22, 4, 17, 8, 9, 14, 10, 6, 10, 8, 9, 10, 9, 21, 4, 8, 9, 4, 25, 8, 6, 11, 4, 15, 5, 19, 6, 4, 7, 9, 14, 4, 13, 10, 21, 11, 6, 4, 17, 8, 9, 6, 5, 37, 6, 4, 10, 9, 4, 7, 15, 15, 4, 15, 7, 22, 5, 13, 12, 26, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                [4, 32, 11, 5, 4, 45, 16, 10, 17, 28, 4, 25, 13, 8, 20, 9, 4, 19, 8, 37, 4, 46, 16, 18, 24, 12, 4, 8, 27, 5, 13, 4, 6, 11, 5, 4, 15, 7, 57, 22, 4, 14, 8, 21, 26, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+            ],
+            'attention_mask': [
+                [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+            ]
+        }
+        # fmt: on
+
+        self.tokenizer_integration_test_util(
+            expected_encoding=expected_encoding,
+            model_name="microsoft/speecht5_asr",
+            revision="c5ef64c71905caeccde0e4462ef3f9077224c524",
+            sequences=sequences,
+        )
+
+    def test_encode_decode(self):
+        tokenizer = SpeechT5Tokenizer.from_pretrained("microsoft/speecht5_tts")
+
+        tokens = tokenizer.tokenize("a = b")
+        self.assertEqual(tokens, ["▁", "a", "▁", "=", "▁", "b"])
+
+        # the `'='` is unknown.
+        ids = tokenizer.convert_tokens_to_ids(tokens)
+        self.assertEqual(ids, [4, 7, 4, 3, 4, 25])
+
+        # let's make sure decoding with the special unknown tokens preserves spaces
+        ids = tokenizer.encode("a = b")
+        self.assertEqual(tokenizer.decode(ids), "a <unk> b</s>")