blackhole33/wev2vec-commonVoice_v1

Author

Mamayusupov Rifat.

Usage

from transformers import SeamlessM4TFeatureExtractor, Wav2Vec2BertProcessor, Wav2Vec2CTCTokenizer, Wav2Vec2BertForCTC
from transformers import pipeline

# Initialize tokenizer
tokenizer = Wav2Vec2CTCTokenizer.from_pretrained("/home/rifat/asr", unk_token="[UNK]", pad_token="[PAD]", word_delimiter_token="|")

# Initialize feature extractor
feature_extractor = SeamlessM4TFeatureExtractor(feature_size=80, num_mel_bins=80, sampling_rate=16000, padding_value=0.0)

# Initialize processor
processor = Wav2Vec2BertProcessor(feature_extractor=feature_extractor, tokenizer=tokenizer)

# Initialize model
model = Wav2Vec2BertForCTC.from_pretrained(
    args.pretrained_model,
    attention_dropout=0.0,
    hidden_dropout=0.0,
    feat_proj_dropout=0.0,
    mask_time_prob=0.0,
    layerdrop=0.0,
    ctc_loss_reduction="mean",
    add_adapter=True,
    pad_token_id=processor.tokenizer.pad_token_id,
    vocab_size=len(processor.tokenizer),
    ignore_mismatched_sizes=True
)

model.config.ctc_zero_infinity = True
model.to("cuda")

# Perform inference
# Initialize the pipeline
pipe = pipeline(model=model, tokenizer=processor.tokenizer, feature_extractor=feature_extractor, task="automatic-speech-recognition")

input_audio = ""
print(pipe(input_audio)['result_text'])