Add wav2vec2 ASR English pretrained model from voxpopuli (#1956)

docs/source/pipelines.rst

@@ -161,6 +161,14 @@ VOXPOPULI_ASR_BASE_10K_DE
    .. autodata:: VOXPOPULI_ASR_BASE_10K_DE
       :no-value:
 
+VOXPOPULI_ASR_BASE_10K_EN
+~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. container:: py attribute
+
+   .. autodata:: VOXPOPULI_ASR_BASE_10K_EN
+      :no-value:
+
 VOXPOPULI_ASR_BASE_10K_ES
 ~~~~~~~~~~~~~~~~~~~~~~~~~
 

test/integration_tests/conftest.py

@@ -35,6 +35,7 @@ def ctc_decoder():
 _FILES = {
     'en': 'Lab41-SRI-VOiCES-src-sp0307-ch127535-sg0042.flac',
     'de': '20090505-0900-PLENARY-16-de_20090505-21_56_00_8.flac',
+    'en2': '20120613-0900-PLENARY-8-en_20120613-13_46_50_3.flac',
     'es': '20130207-0900-PLENARY-7-es_20130207-13_02_05_5.flac',
     'fr': '20121212-0900-PLENARY-5-fr_20121212-11_37_04_10.flac',
     'it': '20170516-0900-PLENARY-16-it_20170516-18_56_31_1.flac',

test/integration_tests/wav2vec2_pipeline_test.py

@@ -18,6 +18,7 @@ from torchaudio.pipelines import (
     HUBERT_XLARGE,
     HUBERT_ASR_LARGE,
     HUBERT_ASR_XLARGE,
+    VOXPOPULI_ASR_BASE_10K_EN,
     VOXPOPULI_ASR_BASE_10K_ES,
     VOXPOPULI_ASR_BASE_10K_DE,
     VOXPOPULI_ASR_BASE_10K_FR,
@@ -57,6 +58,7 @@ def test_pretraining_models(bundle):
         (WAV2VEC2_ASR_LARGE_LV60K_960H, 'en', 'I|HAVE|THAT|CURIOSITY|BESIDE|ME|AT|THIS|MOMENT|'),
         (HUBERT_ASR_LARGE, 'en', 'I|HAVE|THAT|CURIOSITY|BESIDE|ME|AT|THIS|MOMENT|'),
         (HUBERT_ASR_XLARGE, 'en', 'I|HAVE|THAT|CURIOSITY|BESIDE|ME|AT|THIS|MOMENT|'),
+        (VOXPOPULI_ASR_BASE_10K_EN, 'en2', 'i|hope|that|we|will|see|a|ddrasstic|decrease|of|funding|for|the|failed|eu|project|and|that|more|money|will|come|back|to|the|taxpayers'),  # noqa: E501
         (VOXPOPULI_ASR_BASE_10K_ES, 'es', "la|primera|que|es|imprescindible|pensar|a|pequeña|a|escala|para|implicar|y|complementar|así|la|actuación|global"),  # noqa: E501
         (VOXPOPULI_ASR_BASE_10K_DE, 'de', "dabei|spielt|auch|eine|sorgfältige|berichterstattung|eine|wichtige|rolle"),
         (VOXPOPULI_ASR_BASE_10K_FR, 'fr', 'la|commission|va|faire|des|propositions|sur|ce|sujet|comment|mettre|en|place|cette|capacité|fiscale|et|le|conseil|européen|y|reviendra|sour|les|sujets|au|moins|de|mars'),  # noqa: E501

torchaudio/pipelines/__init__.py

@@ -14,6 +14,7 @@ from ._wav2vec2.impl import (
     WAV2VEC2_ASR_LARGE_LV60K_100H,
     WAV2VEC2_ASR_LARGE_LV60K_960H,
     WAV2VEC2_XLSR53,
+    VOXPOPULI_ASR_BASE_10K_EN,
     VOXPOPULI_ASR_BASE_10K_ES,
     VOXPOPULI_ASR_BASE_10K_DE,
     VOXPOPULI_ASR_BASE_10K_FR,
@@ -48,6 +49,7 @@ __all__ = [
     'WAV2VEC2_ASR_LARGE_LV60K_100H',
     'WAV2VEC2_ASR_LARGE_LV60K_960H',
     'WAV2VEC2_XLSR53',
+    'VOXPOPULI_ASR_BASE_10K_EN',
     'VOXPOPULI_ASR_BASE_10K_ES',
     'VOXPOPULI_ASR_BASE_10K_DE',
     'VOXPOPULI_ASR_BASE_10K_FR',

torchaudio/pipelines/_wav2vec2/impl.py

@@ -1033,6 +1033,50 @@ Please refer to :func:`torchaudio.pipelines.Wav2Vec2ASRBundle` for the usage.
 """  # noqa: E501
 
 
+VOXPOPULI_ASR_BASE_10K_EN = Wav2Vec2ASRBundle(
+    'wav2vec2_voxpopuli_base_10k_asr_en.pt',
+    {
+        "extractor_mode": "group_norm",
+        "extractor_conv_layer_config": [
+            (512, 10, 5),
+            (512, 3, 2),
+            (512, 3, 2),
+            (512, 3, 2),
+            (512, 3, 2),
+            (512, 2, 2),
+            (512, 2, 2),
+        ],
+        "extractor_conv_bias": False,
+        "encoder_embed_dim": 768,
+        "encoder_projection_dropout": 0.0,
+        "encoder_pos_conv_kernel": 128,
+        "encoder_pos_conv_groups": 16,
+        "encoder_num_layers": 12,
+        "encoder_num_heads": 12,
+        "encoder_attention_dropout": 0.0,
+        "encoder_ff_interm_features": 3072,
+        "encoder_ff_interm_dropout": 0.1,
+        "encoder_dropout": 0.0,
+        "encoder_layer_norm_first": False,
+        "encoder_layer_drop": 0.1,
+        "aux_num_out": 28
+    },
+    _labels=utils._get_vp_en_labels(),
+    _sample_rate=16000,
+    _remove_aux_axis=(1, 2, 3, 31),
+)
+VOXPOPULI_ASR_BASE_10K_EN.__doc__ = """wav2vec 2.0 model with "Base" configuration.
+Pre-trained on 10k hours of unlabeled audio from *VoxPopuli* dataset [:footcite:`voxpopuli`]
+("10k" subset, consisting of 23 languages).
+Fine-tuned for ASR on 543 hours of transcribed audio from "en" subset.
+Originally published by the authors of *VoxPopuli* [:footcite:`voxpopuli`] under CC BY-NC 4.0 and
+redistributed with the same license.
+[`License <https://github.com/facebookresearch/voxpopuli/tree/160e4d7915bad9f99b2c35b1d3833e51fd30abf2#license>`__,
+`Source <https://github.com/facebookresearch/voxpopuli/tree/160e4d7915bad9f99b2c35b1d3833e51fd30abf2#asr-and-lm>`__]
+Please refer to :func:`torchaudio.pipelines.Wav2Vec2ASRBundle` for the usage.
+"""  # noqa: E501
+
+
 VOXPOPULI_ASR_BASE_10K_ES = Wav2Vec2ASRBundle(
     'wav2vec2_voxpopuli_base_10k_asr_es.pt',
     {

torchaudio/pipelines/_wav2vec2/utils.py

@@ -67,6 +67,38 @@ def _get_de_labels():
     )
 
 
+def _get_vp_en_labels():
+    return (
+        "|",
+        "e",
+        "t",
+        "o",
+        "i",
+        "a",
+        "n",
+        "s",
+        "r",
+        "h",
+        "l",
+        "d",
+        "c",
+        "u",
+        "m",
+        "p",
+        "f",
+        "g",
+        "w",
+        "y",
+        "b",
+        "v",
+        "k",
+        "x",
+        "j",
+        "q",
+        "z",
+    )
+
+
 def _get_es_labels():
     return (
         "|",