fix text featurizer

.notebook/jit_infer.ipynb

@@ -295,6 +295,7 @@
    "source": [
     "dataset = ManifestDataset(\n",
     "        config.data.test_manifest,\n",
+    "        config.data.unit_type,\n",
     "        config.data.vocab_filepath,\n",
     "        config.data.mean_std_filepath,\n",
     "        augmentation_config=\"{}\",\n",

deepspeech/exps/deepspeech2/bin/deploy/runtime.py

@@ -85,8 +85,10 @@ def start_server(config, args):
     """Start the ASR server"""
     dataset = ManifestDataset(
         config.data.test_manifest,
+        config.data.unit_type,
         config.data.vocab_filepath,
         config.data.mean_std_filepath,
+        spm_model_prefix=config.data.spm_model_prefix,
         augmentation_config="{}",
         max_duration=config.data.max_duration,
         min_duration=config.data.min_duration,

deepspeech/exps/deepspeech2/bin/deploy/server.py

@@ -37,8 +37,10 @@ def start_server(config, args):
     """Start the ASR server"""
     dataset = ManifestDataset(
         config.data.test_manifest,
+        config.data.unit_type,
         config.data.vocab_filepath,
         config.data.mean_std_filepath,
+        spm_model_prefix=config.data.spm_model_prefix,
         augmentation_config="{}",
         max_duration=config.data.max_duration,
         min_duration=config.data.min_duration,

deepspeech/exps/deepspeech2/bin/tune.py

@@ -43,8 +43,10 @@ def tune(config, args):
 
     dev_dataset = ManifestDataset(
         config.data.dev_manifest,
+        config.data.unit_type,
         config.data.vocab_filepath,
         config.data.mean_std_filepath,
+        spm_model_prefix=config.data.spm_model_prefix,
         augmentation_config="{}",
         max_duration=config.data.max_duration,
         min_duration=config.data.min_duration,

deepspeech/exps/deepspeech2/config.py

@@ -21,7 +21,9 @@ _C.data = CN(
         train_manifest="",
         dev_manifest="",
         test_manifest="",
+        unit_type="char",
         vocab_filepath="",
+        spm_model_prefix="",
         mean_std_filepath="",
         augmentation_config="",
         max_duration=float('inf'),

deepspeech/exps/deepspeech2/model.py

@@ -148,8 +148,10 @@ class DeepSpeech2Trainer(Trainer):
 
         train_dataset = ManifestDataset(
             config.data.train_manifest,
+            config.data.unit_type,
             config.data.vocab_filepath,
             config.data.mean_std_filepath,
+            spm_model_prefix=config.data.spm_model_prefix,
             augmentation_config=io.open(
                 config.data.augmentation_config, mode='r',
                 encoding='utf8').read(),
@@ -168,8 +170,10 @@ class DeepSpeech2Trainer(Trainer):
 
         dev_dataset = ManifestDataset(
             config.data.dev_manifest,
+            config.data.unit_type,
             config.data.vocab_filepath,
             config.data.mean_std_filepath,
+            spm_model_prefix=config.data.spm_model_prefix,
             augmentation_config="{}",
             max_duration=config.data.max_duration,
             min_duration=config.data.min_duration,
@@ -361,8 +365,10 @@ class DeepSpeech2Tester(DeepSpeech2Trainer):
         # return raw text
         test_dataset = ManifestDataset(
             config.data.test_manifest,
+            config.data.unit_type,
             config.data.vocab_filepath,
             config.data.mean_std_filepath,
+            spm_model_prefix=config.data.spm_model_prefix,
             augmentation_config="{}",
             max_duration=config.data.max_duration,
             min_duration=config.data.min_duration,

deepspeech/frontend/featurizer/audio_featurizer.py

@@ -109,7 +109,7 @@ class AudioFeaturizer(object):
             feat_dim = int(fft_point * (self._target_sample_rate / 1000) / 2 +
                            1)
         elif self._specgram_type == 'mfcc':
-            # mfcc,delta, delta-delta
+            # mfcc, delta, delta-delta
             feat_dim = int(13 * 3)
         else:
             raise ValueError("Unknown specgram_type %s. "

deepspeech/frontend/featurizer/speech_featurizer.py

@@ -52,7 +52,9 @@ class SpeechFeaturizer(object):
     """
 
     def __init__(self,
+                 unit_type,
                  vocab_filepath,
+                 spm_model_prefix=None,
                  specgram_type='linear',
                  stride_ms=10.0,
                  window_ms=20.0,
@@ -70,7 +72,8 @@ class SpeechFeaturizer(object):
             target_sample_rate=target_sample_rate,
             use_dB_normalization=use_dB_normalization,
             target_dB=target_dB)
-        self._text_featurizer = TextFeaturizer(vocab_filepath)
+        self._text_featurizer = TextFeaturizer(unit_type, vocab_filepath,
+                                               spm_model_prefix)
 
     def featurize(self, speech_segment, keep_transcription_text):
         """Extract features for speech segment.
@@ -116,4 +119,4 @@ class SpeechFeaturizer(object):
         :return: audio feature size.
         :rtype: int
         """
-        return self._audio_featurizer.feature_size
+        return self._audio_featurizer.feature_size

deepspeech/frontend/featurizer/text_featurizer.py

@@ -15,25 +15,35 @@
 
 import os
 import codecs
+import sentencepiece as spm
 
+from deepspeech.frontend.utility import UNK
 
-class TextFeaturizer(object):
-    """Text featurizer, for processing or extracting features from text.
 
-    Currently, it only supports char-level tokenizing and conversion into
+class TextFeaturizer(object):
-    a list of token indices. Note that the token indexing order follows the
+    def __init__(self, unit_type, vocab_filepath, spm_model_prefix=None):
-    given vocabulary file.
+        """Text featurizer, for processing or extracting features from text.
 
-    :param vocab_filepath: Filepath to load vocabulary for token indices
+        Currently, it supports char/word/sentence-piece level tokenizing and conversion into
-                           conversion.
+        a list of token indices. Note that the token indexing order follows the
-    :type specgram_type: str
+        given vocabulary file.
-    """
 
-    def __init__(self, vocab_filepath):
+        Args:
-        self.unk = '<unk>'
+            unit_type (str): unit type, e.g. char, word, spm
+            vocab_filepath (str): Filepath to load vocabulary for token indices conversion.
+            spm_model_prefix (str, optional): spm model prefix. Defaults to None.
+        """
+        assert unit_type in ('char', 'spm', 'word')
+        self.unk = UNK
+        self.unit_type = unit_type
         self._vocab_dict, self._vocab_list = self._load_vocabulary_from_file(
             vocab_filepath)
 
+        if unit_type == 'spm':
+            spm_model = spm_model_prefix + '.model'
+            self.sp = spm.SentencePieceProcessor()
+            self.sp.Load(self.spm_model)
+
     def featurize(self, text):
         """Convert text string to a list of token indices in char-level.Note
         that the token indexing order follows the given vocabulary file.
@@ -43,7 +53,13 @@ class TextFeaturizer(object):
         :return: List of char-level token indices.
         :rtype: list
         """
-        tokens = self._char_tokenize(text)
+        if unit_type == 'char':
+            tokens = self._char_tokenize(text)
+        elif unit_type == 'word':
+            tokens = self._word_tokenize(text)
+        else:
+            tokens = self._spm_tokenize(text)
+
         ids = []
         for token in tokens:
             token = token if token in self._vocab_dict else self.unk
@@ -72,6 +88,42 @@ class TextFeaturizer(object):
         """Character tokenizer."""
         return list(text.strip())
 
+    def _word_tokenize(self, text):
+        """Word tokenizer, spearte by <space>."""
+        return text.strip().split()
+
+    def _spm_tokenize(self, text):
+        """spm tokenize.
+
+        Args:
+            text (str): text string.
+
+        Returns:
+            List[str]: sentence pieces str code
+        """
+        stats = {"num_empty": 0, "num_filtered": 0}
+
+        def valid(line):
+            return True
+
+        def encode(l):
+            return self.sp.EncodeAsPieces(l)
+
+        def encode_line(line):
+            line = line.strip()
+            if len(line) > 0:
+                line = encode(line)
+                if valid(line):
+                    return line
+                else:
+                    stats["num_filtered"] += 1
+            else:
+                stats["num_empty"] += 1
+            return None
+
+        enc_line = encode_line(text)
+        return enc_line
+
     def _load_vocabulary_from_file(self, vocab_filepath):
         """Load vocabulary from file."""
         vocab_lines = []

deepspeech/frontend/normalizer.py

@@ -16,6 +16,7 @@
 import numpy as np
 import random
 from deepspeech.frontend.utility import read_manifest
+from deepspeech.frontend.utility import load_cmvn
 from deepspeech.frontend.audio import AudioSegment
 
 
@@ -79,10 +80,8 @@ class FeatureNormalizer(object):
 
     def _read_mean_std_from_file(self, filepath, eps=1e-20):
         """Load mean and std from file."""
-        npzfile = np.load(filepath)
+        mean, std = load_cmvn(filepath, filetype='npz')
-        self._mean = npzfile["mean"]
+        self._mean = mean
-        std = npzfile["std"]
-        std = np.clip(std, eps, None)
         self._istd = 1.0 / std
 
     def _compute_mean_std(self, manifest_path, featurize_func, num_samples):
@@ -92,8 +91,7 @@ class FeatureNormalizer(object):
         features = []
         for instance in sampled_manifest:
             features.append(
-                featurize_func(
+                featurize_func(AudioSegment.from_file(instance["feat"])))
-                    AudioSegment.from_file(instance["audio_filepath"])))
         features = np.hstack(features)  #(D, T)
         self._mean = np.mean(features, axis=1).reshape([-1, 1])  #(D, 1)
         self._std = np.std(features, axis=1).reshape([-1, 1])  #(D, 1)

deepspeech/frontend/utility.py

@@ -20,6 +20,7 @@ import os
 import tarfile
 import time
 import logging
+from typing import List
 from threading import Thread
 from multiprocessing import Process, Manager, Value
 
@@ -39,31 +40,32 @@ EOS = SOS
 UNK = "<unk>"
 BLANK = "<blank>"
 
-# """Load and parse manifest file.
-
-# Instances with durations outside [min_duration, max_duration] will be
-# filtered out.
-
-# :param manifest_path: Manifest file to load and parse.
-# :type manifest_path: str
-# :param max_duration:maximum output seq length, in seconds for raw wav, in frame numbers for feature data.
-# :type max_duration: float
-# :param min_duration: minimum input seq length, in seconds for raw wav, in frame numbers for feature data.
-# :type min_duration: float
-# :return: Manifest parsing results. List of dict.
-# :rtype: list
-# :raises IOError: If failed to parse the manifest.
-# """
-
 
 def read_manifest(
         manifest_path,
         max_input_len=float('inf'),
         min_input_len=0.0,
-        max_output_len=500.0,
+        max_output_len=float('inf'),
         min_output_len=0.0,
-        max_output_input_ratio=10.0,
+        max_output_input_ratio=float('inf'),
-        min_output_input_ratio=0.05, ):
+        min_output_input_ratio=0.0, ):
+    """Load and parse manifest file.
+
+    Args:
+        manifest_path ([type]): Manifest file to load and parse.
+        max_input_len ([type], optional): maximum output seq length, in seconds for raw wav, in frame numbers for feature data. Defaults to float('inf').
+        min_input_len (float, optional): minimum input seq length, in seconds for raw wav, in frame numbers for feature data. Defaults to 0.0.
+        max_output_len (float, optional): maximum input seq length, in modeling units. Defaults to 500.0.
+        min_output_len (float, optional): minimum input seq length, in modeling units. Defaults to 0.0.
+        max_output_input_ratio (float, optional): maximum output seq length/output seq length ratio. Defaults to 10.0.
+        min_output_input_ratio (float, optional): minimum output seq length/output seq length ratio. Defaults to 0.05.
+
+    Raises:
+        IOError: If failed to parse the manifest.
+
+    Returns:
+        List[dict]: Manifest parsing results.
+    """
 
     manifest = []
     for json_line in codecs.open(manifest_path, 'r', 'utf-8'):
@@ -71,33 +73,23 @@ def read_manifest(
             json_data = json.loads(json_line)
         except Exception as e:
             raise IOError("Error reading manifest: %s" % str(e))
-        feat_len = json_data["feat_shape"][0]
+
-        token_len = json_data["token_shape"][0]
+        feat_len = json_data["feat_shape"][
+            0] if 'feat_shape' in json_data else 1.0
+        token_len = json_data["token_shape"][
+            0] if 'token_shape' in json_data else 1.0
         conditions = [
-            feat_len > min_input_len,
+            feat_len >= min_input_len,
-            feat_len < max_input_len,
+            feat_len <= max_input_len,
-            token_len > min_output_len,
+            token_len >= min_output_len,
-            token_len < max_output_len,
+            token_len <= max_output_len,
-            token_len / feat_len > min_output_input_ratio,
+            token_len / feat_len >= min_output_input_ratio,
-            token_len / feat_len < max_output_input_ratio,
+            token_len / feat_len <= max_output_input_ratio,
         ]
         if all(conditions):
             manifest.append(json_data)
     return manifest
 
-    # parser.add_argument('--max_input_len', type=float,
-    #                     default=20,
-    #                     help='maximum output seq length, in seconds for raw wav, in frame numbers for feature data')
-    # parser.add_argument('--min_output_len', type=float,
-    #                     default=0, help='minimum input seq length, in modeling units')
-    # parser.add_argument('--max_output_len', type=float,
-    #                     default=500,
-    #                     help='maximum output seq length, in modeling units')
-    # parser.add_argument('--min_output_input_ratio', type=float, default=0.05,
-    #                     help='minimum output seq length/output seq length ratio')
-    # parser.add_argument('--max_output_input_ratio', type=float, default=10,
-    #                     help='maximum output seq length/output seq length ratio')
-
 
 def rms_to_db(rms: float):
     """Root Mean Square to dB.
@@ -251,8 +243,8 @@ def _load_kaldi_cmvn(kaldi_cmvn_file):
 
 def _load_npz_cmvn(npz_cmvn_file, eps=1e-20):
     npzfile = np.load(npz_cmvn_file)
-    means = npzfile["mean"]
+    means = npzfile["mean"]  #(D, 1)
-    std = npzfile["std"]
+    std = npzfile["std"]  #(D, 1)
     std = np.clip(std, eps, None)
     variance = 1.0 / std
     cmvn = np.array([means, variance])
@@ -278,7 +270,7 @@ def load_cmvn(cmvn_file: str, filetype: str):
         cmvn = _load_json_cmvn(cmvn_file)
     elif filetype == "kaldi":
         cmvn = _load_kaldi_cmvn(cmvn_file)
-    elif filtype == "npz":
+    elif filetype == "npz":
         cmvn = _load_npz_cmvn(cmvn_file)
     else:
         raise ValueError(f"cmvn file type no support: {filetype}")

deepspeech/io/__init__.py

@@ -21,8 +21,10 @@ from deepspeech.io.dataset import ManifestDataset
 
 
 def create_dataloader(manifest_path,
+                      unit_type,
                       vocab_filepath,
                       mean_std_filepath,
+                      spm_model_prefix,
                       augmentation_config='{}',
                       max_duration=float('inf'),
                       min_duration=0.0,
@@ -42,8 +44,10 @@ def create_dataloader(manifest_path,
 
     dataset = ManifestDataset(
         manifest_path,
+        unit_type,
         vocab_filepath,
         mean_std_filepath,
+        spm_model_prefix=spm_model_prefix,
         augmentation_config=augmentation_config,
         max_duration=max_duration,
         min_duration=min_duration,

deepspeech/io/dataset.py

@@ -38,8 +38,10 @@ __all__ = [
 class ManifestDataset(Dataset):
     def __init__(self,
                  manifest_path,
+                 unit_type,
                  vocab_filepath,
                  mean_std_filepath,
+                 spm_model_prefix=None,
                  augmentation_config='{}',
                  max_duration=float('inf'),
                  min_duration=0.0,
@@ -57,8 +59,10 @@ class ManifestDataset(Dataset):
 
         Args:
             manifest_path (str): manifest josn file path
-            vocab_filepath (str): vocab file path
+            unit_type(str): token unit type, e.g. char, word, spm
+            vocab_filepath (str): vocab file path.
             mean_std_filepath (str): mean and std file path, which suffix is *.npy
+            spm_model_prefix (str): spm model prefix, need if `unit_type` is spm.
             augmentation_config (str, optional): augmentation json str. Defaults to '{}'.
             max_duration (float, optional): audio length in seconds must less than this. Defaults to float('inf').
             min_duration (float, optional): audio length is seconds must greater than this. Defaults to 0.0.
@@ -78,10 +82,12 @@ class ManifestDataset(Dataset):
         self._max_duration = max_duration
         self._min_duration = min_duration
         self._normalizer = FeatureNormalizer(mean_std_filepath)
-        self._augmentation_pipeline = AugmentationPipeline(
+        self._audio_augmentation_pipeline = AugmentationPipeline(
             augmentation_config=augmentation_config, random_seed=random_seed)
         self._speech_featurizer = SpeechFeaturizer(
+            unit_type=unit_type,
             vocab_filepath=vocab_filepath,
+            spm_model_prefix=spm_model_prefix,
             specgram_type=specgram_type,
             stride_ms=stride_ms,
             window_ms=window_ms,
@@ -174,7 +180,7 @@ class ManifestDataset(Dataset):
                 self._subfile_from_tar(audio_file), transcript)
         else:
             speech_segment = SpeechSegment.from_file(audio_file, transcript)
-        self._augmentation_pipeline.transform_audio(speech_segment)
+        self._audio_augmentation_pipeline.transform_audio(speech_segment)
         specgram, transcript_part = self._speech_featurizer.featurize(
             speech_segment, self._keep_transcription_text)
         specgram = self._normalizer.apply(specgram)
@@ -191,7 +197,7 @@ class ManifestDataset(Dataset):
 
         def reader():
             for instance in manifest:
-                inst = self.process_utterance(instance["audio_filepath"],
+                inst = self.process_utterance(instance["feat"],
                                               instance["text"])
                 yield inst
 
@@ -202,5 +208,4 @@ class ManifestDataset(Dataset):
 
     def __getitem__(self, idx):
         instance = self._manifest[idx]
-        return self.process_utterance(instance["audio_filepath"],
+        return self.process_utterance(instance["feat"], instance["text"])
-                                      instance["text"])

examples/tiny/s0/local/data.sh

@@ -52,6 +52,7 @@ fi
 # format manifest with tokenids, vocab size
 python3 ${MAIN_ROOT}/utils/format_data.py \
 --feat_type "raw" \
+--cmvn_path "data/mean_std.npz" \
 --unit_type "bpe" \
 --bpe_model_prefix ${bpeprefix} \
 --vocab_path="data/vocab.txt" \

requirements.txt

@@ -3,6 +3,7 @@ resampy==0.2.2
 SoundFile==0.9.0.post1
 python_speech_features
 tensorboardX
+sentencepiece
 yacs
 typeguard
 pre-commit

utils/format_data.py

@@ -24,6 +24,7 @@ from deepspeech.frontend.utility import read_manifest
 from deepspeech.frontend.utility import UNK
 from deepspeech.frontend.utility import BLANK
 from deepspeech.frontend.utility import SOS
+from deepspeech.frontend.utility import load_cmvn
 from deepspeech.utils.utility import add_arguments
 from deepspeech.utils.utility import print_arguments
 
@@ -31,10 +32,13 @@ parser = argparse.ArgumentParser(description=__doc__)
 add_arg = functools.partial(add_arguments, argparser=parser)
 # yapf: disable
 add_arg('feat_type', str, "raw", "speech feature type, e.g. raw(wav, flac), kaldi")
+add_arg('cmvn_path',       str,
+        'examples/librispeech/data/mean_std.npz',
+        "Filepath of cmvn.")
 add_arg('unit_type', str, "character", "Unit type, e.g. character, word, bpe")
 add_arg('vocab_path',       str,
         'examples/librispeech/data/vocab.txt',
-        "Filepath to write the vocabulary.")
+        "Filepath of the vocabulary.")
 add_arg('manifest_paths',   str,
         None,
         "Filepaths of manifests for building vocabulary. "
@@ -51,6 +55,11 @@ args = parser.parse_args()
 def main():
     print_arguments(args)
 
+    # get feat dim
+    mean, std = load_cmvn(args.cmvn_path, filetype='npz')
+    feat_dim = mean.shape[0]
+    print(f"Feature dim: {feat_dim}")
+
     # read vocab
     vocab = dict()
     with open(args.vocab_path, 'r', encoding='utf-8') as fin:
@@ -58,6 +67,7 @@ def main():
             token = line.strip()
             vocab[token] = len(vocab)
     vocab_size = len(vocab)
+    print(f"Vocab size: {vocab_size}")
 
     fout = open(args.output_path, 'w', encoding='utf-8')
 
@@ -78,6 +88,12 @@ def main():
                 line_json['token'] = tokens
                 line_json['token_id'] = tokenids
                 line_json['token_shape'] = (len(tokenids), vocab_size)
+                feat_shape = line_json['feat_shape']
+                assert isinstance(feat_shape, (list, tuple)), type(feat_shape)
+                if args.feat_type == 'raw':
+                    feat_shape.append(feat_dim)
+                else: # kaldi
+                    raise NotImplemented('no support kaldi feat now!')
                 fout.write(json.dumps(line_json) + '\n')
     else:
         import sentencepiece as spm
@@ -118,6 +134,12 @@ def main():
                 line_json['token'] = tokens
                 line_json['token_id'] = tokenids
                 line_json['token_shape'] = (len(tokenids), vocab_size)
+                feat_shape = line_json['feat_shape']
+                assert isinstance(feat_shape, (list, tuple)), type(feat_shape)
+                if args.feat_type == 'raw':
+                    feat_shape.append(feat_dim)
+                else: # kaldi
+                    raise NotImplemented('no support kaldi feat now!')
                 fout.write(json.dumps(line_json) + '\n')
 
     fout.close()