def __init__(self, task, wer_args, zero_infinity, sentence_avg,
remove_bpe):
super().__init__(task)
self.blank_idx = task.target_dictionary.bos()
self.pad_idx = task.target_dictionary.pad()
self.eos_idx = task.target_dictionary.eos()
self.post_process = remove_bpe if remove_bpe else "letter"
if wer_args is not None:
from examples.speech_recognition.w2l_decoder import W2lKenLMDecoder
wer_compute_kenlm, wer_lexicon, lm_w, ws_w = eval(wer_args)
dec_args = Namespace()
dec_args.nbest = 1
dec_args.criterion = "ctc"
dec_args.kenlm_model = wer_compute_kenlm
dec_args.lexicon = wer_lexicon
dec_args.beam = 50
dec_args.beam_size_token = min(50, len(task.target_dictionary))
dec_args.beam_threshold = min(50, len(task.target_dictionary))
dec_args.lm_weight = lm_w
dec_args.word_score = ws_w
dec_args.unk_weight = -math.inf
dec_args.sil_weight = 0
self.w2l_decoder = W2lKenLMDecoder(dec_args,
task.target_dictionary)
else:
self.w2l_decoder = None
self.zero_infinity = zero_infinity
self.sentence_avg = sentence_avg
def build_generator(cfg: UnsupGenerateConfig):
w2l_decoder = cfg.w2l_decoder
if w2l_decoder == DecoderType.VITERBI:
from examples.speech_recognition.w2l_decoder import W2lViterbiDecoder
return W2lViterbiDecoder(cfg, task.target_dictionary)
elif w2l_decoder == DecoderType.KENLM:
from examples.speech_recognition.w2l_decoder import W2lKenLMDecoder
return W2lKenLMDecoder(cfg, task.target_dictionary)
elif w2l_decoder == DecoderType.FAIRSEQ:
from examples.speech_recognition.w2l_decoder import W2lFairseqLMDecoder
return W2lFairseqLMDecoder(cfg, task.target_dictionary)
elif w2l_decoder == DecoderType.KALDI:
from examples.speech_recognition.kaldi.kaldi_decoder import KaldiDecoder
assert cfg.kaldi_decoder_config is not None
return KaldiDecoder(
cfg.kaldi_decoder_config,
cfg.beam,
)
else:
raise NotImplementedError(
"only wav2letter decoders with (viterbi, kenlm, fairseqlm) options are supported at the moment but found "
+ str(w2l_decoder)
)
def build_generator(self, args):
w2l_decoder = getattr(args, "w2l_decoder", None)
if w2l_decoder == "viterbi":
from examples.speech_recognition.w2l_decoder import W2lViterbiDecoder
return W2lViterbiDecoder(args, self.target_dictionary)
elif w2l_decoder == "kenlm":
from examples.speech_recognition.w2l_decoder import W2lKenLMDecoder
return W2lKenLMDecoder(args, self.target_dictionary)
else:
return super().build_generator(args)
def get_decoder(decoder_args_dict, dictionary):
decoder_args = Namespace(**decoder_args_dict)
if decoder_args.decoder_type == "viterbi":
from examples.speech_recognition.w2l_decoder import W2lViterbiDecoder
return W2lViterbiDecoder(decoder_args, dictionary)
elif decoder_args.decoder_type == "kenlm":
from examples.speech_recognition.w2l_decoder import W2lKenLMDecoder
decoder_args.beam_size_token = len(dictionary)
if isinstance(decoder_args.unk_weight, str):
decoder_args.unk_weight = eval(decoder_args.unk_weight)
return W2lKenLMDecoder(decoder_args, dictionary)
return None
def build_generator(args):
w2l_decoder = getattr(args, "w2l_decoder", None)
if w2l_decoder == "viterbi":
from examples.speech_recognition.w2l_decoder import W2lViterbiDecoder
return W2lViterbiDecoder(args, task.target_dictionary)
elif w2l_decoder == "kenlm":
from examples.speech_recognition.w2l_decoder import W2lKenLMDecoder
print(task.target_dictionary.symbols)
return W2lKenLMDecoder(args, task.target_dictionary)
elif w2l_decoder == "fairseqlm":
from examples.speech_recognition.w2l_decoder import W2lFairseqLMDecoder
print(task.target_dictionary.symbols)
return W2lFairseqLMDecoder(args, task.target_dictionary)
else:
return super().build_generator(args)
def build_generator(args):
w2l_decoder = getattr(args, "w2l_decoder", None)
if w2l_decoder == "viterbi":
from examples.speech_recognition.w2l_decoder import W2lViterbiDecoder
return W2lViterbiDecoder(args, task.target_dictionary)
elif w2l_decoder == "kenlm":
from examples.speech_recognition.w2l_decoder import W2lKenLMDecoder
return W2lKenLMDecoder(args, task.target_dictionary)
elif w2l_decoder == "fairseqlm":
from examples.speech_recognition.w2l_decoder import W2lFairseqLMDecoder
return W2lFairseqLMDecoder(args, task.target_dictionary)
else:
print('only wav2letter decoders with (viterbi, kenlm, fairseqlm) options are supported at the moment')
def __init__(self, cfg: CtcCriterionConfig, task: FairseqTask):
super().__init__(task)
self.blank_idx = (
task.target_dictionary.index(task.blank_symbol)
if hasattr(task, "blank_symbol")
else 0
)
self.pad_idx = task.target_dictionary.pad()
self.eos_idx = task.target_dictionary.eos()
self.post_process = cfg.post_process
if cfg.wer_args is not None:
(
cfg.wer_kenlm_model,
cfg.wer_lexicon,
cfg.wer_lm_weight,
cfg.wer_word_score,
) = eval(cfg.wer_args)
if cfg.wer_kenlm_model is not None and cfg.wer_kenlm_model != "":
from examples.speech_recognition.w2l_decoder import W2lKenLMDecoder
dec_args = Namespace()
dec_args.nbest = 1
dec_args.criterion = "ctc"
dec_args.kenlm_model = cfg.wer_kenlm_model
dec_args.lexicon = cfg.wer_lexicon
dec_args.beam = 50
dec_args.beam_size_token = min(50, len(task.target_dictionary))
dec_args.beam_threshold = min(50, len(task.target_dictionary))
dec_args.lm_weight = cfg.wer_lm_weight
dec_args.word_score = cfg.wer_word_score
dec_args.unk_weight = -math.inf
dec_args.sil_weight = 0
self.w2l_decoder = W2lKenLMDecoder(dec_args, task.target_dictionary)
else:
self.w2l_decoder = None
self.zero_infinity = cfg.zero_infinity
self.sentence_avg = cfg.sentence_avg
def build_generator(args):
w2l_decoder = getattr(args, "w2l_decoder", None)
if w2l_decoder == "viterbi":
from examples.speech_recognition.w2l_decoder import W2lViterbiDecoder
return W2lViterbiDecoder(args, task.target_dictionary)
elif w2l_decoder == "kenlm":
from examples.speech_recognition.w2l_decoder import W2lKenLMDecoder
return W2lKenLMDecoder(args, task.target_dictionary)
elif w2l_decoder == "fairseqlm":
from examples.speech_recognition.w2l_decoder import W2lFairseqLMDecoder
return W2lFairseqLMDecoder(args, task.target_dictionary)
elif w2l_decoder == "ctc_decoder":
from examples.speech_recognition.ctc_decoder import CTCDecoder
return CTCDecoder(args, task.target_dictionary)
elif w2l_decoder == "cif_decoder":
from examples.speech_recognition.cif_decoder import CIFDecoder
return CIFDecoder(args, task.target_dictionary, {})
elif w2l_decoder == "cif_lm_decoder":
from examples.speech_recognition.cif_decoder import CIFDecoder
return CIFDecoder(args, task.target_dictionary, ({}, {}))
elif w2l_decoder == "cif_bert_decoder":
from examples.speech_recognition.cif_bert_decoder import CIF_BERT_Decoder
return CIF_BERT_Decoder(args, task.target_dictionary)
elif w2l_decoder == "seq2seq_decoder":
from examples.speech_recognition.seq2seq_decoder import Seq2seqDecoder
return Seq2seqDecoder(args, task.target_dictionary, {})
elif w2l_decoder == "seq2seq_lm_decoder":
from examples.speech_recognition.seq2seq_decoder import Seq2seqDecoder
return Seq2seqDecoder(args, task.target_dictionary, ({}, {}))
else:
return super().build_generator(args)