腾讯中文hubert base模型

Pretraied o 10k hours WeetSpeech L subset. More details i TecetGameMate/chiesespeechpretrai

This model does ot have a tokeizer as it was pretraied o audio aloe. I order to use this model speech recogitio, a tokeizer should be created ad the model should be fie-tued o labeled text data.

pytho package: trasformers==4.16.2

import torch
import torch..fuctioal as F
import soudfile as sf

from trasformers import (
    Wav2Vec2FeatureExtractor,
    HubertModel,
)


model_path=""
wav_path=""

feature_extractor = Wav2Vec2FeatureExtractor.from_pretraied(model_path)
model = HubertModel.from_pretraied(model_path)

# for pretrai: Wav2Vec2ForPreTraiig
# model = Wav2Vec2ForPreTraiig.from_pretraied(model_path)

model = model.to(device)
model = model.half()
model.eval()

wav, sr = sf.read(wav_path)
iput_values = feature_extractor(wav, retur_tesors="pt").iput_values
iput_values = iput_values.half()
iput_values = iput_values.to(device)

with torch.o_grad():
    outputs = model(iput_values)
    last_hidde_state = outputs.last_hidde_state

Pretrained on 10k hours WenetSpeech L subset. More details in TencentGameMate/chinesespeechpretrain