update

app.py

@@ -1,18 +1,15 @@
 import gradio as gr
 import nltk
 import librosa
-from optimum.onnxruntime import ORTModelForSeq2SeqLM
-
-from transformers import pipeline, TranslationPipeline, AutoTokenizer, TranslationPipeline
-from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC, Wav2Vec2Tokenizer
+from transformers import pipeline
 from transformers.file_utils import cached_path, hf_bucket_url
 import os, zipfile
+from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC, Wav2Vec2Tokenizer
 from datasets import load_dataset
 import torch
 import kenlm
 import torchaudio
 from pyctcdecode import Alphabet, BeamSearchDecoderCTC, LanguageModel
-device = torch.device(0 if torch.cuda.is_available() else "cpu")
 
 """Vietnamese speech2text"""
 cache_dir = './cache/'
@@ -126,17 +123,8 @@ def speech2text_en(input_file):
 """Machine translation"""
 vien_model_checkpoint = "datnth1709/finetuned_HelsinkiNLP-opus-mt-vi-en_PhoMT"
 envi_model_checkpoint = "datnth1709/finetuned_HelsinkiNLP-opus-mt-en-vi_PhoMT"
-# vien_translator = pipeline("translation", model=vien_model_checkpoint)
-# envi_translator = pipeline("translation", model=envi_model_checkpoint)
-
-vien_tokenizer = AutoTokenizer.from_pretrained(vien_model_checkpoint, return_tensors="pt")
-vien_model = ORTModelForSeq2SeqLM.from_pretrained(vien_model_checkpoint)
-vien_translator = TranslationPipeline(model=vien_model, tokenizer=vien_tokenizer,clean_up_tokenization_spaces=True, device=device)
-
-envi_tokenizer = AutoTokenizer.from_pretrained(envi_model_checkpoint, return_tensors="pt")
-envi_model = ORTModelForSeq2SeqLM.from_pretrained(envi_model_checkpoint)
-envi_translator = TranslationPipeline(model=envi_model, tokenizer=envi_tokenizer,clean_up_tokenization_spaces=True, device=device)
-
+vien_translator = pipeline("translation", model=vien_model_checkpoint)
+envi_translator = pipeline("translation", model=envi_model_checkpoint)
 
 def translate_vi2en(Vietnamese):
     return vien_translator(Vietnamese)[0]['translation_text']
@@ -188,9 +176,9 @@ def transcribe_en(audio, state_en="", state_vi=""):
     transcription = eng_tokenizer.decode(predicted_ids[0])
     # Output is all upper case
     transcription = correct_casing(transcription.lower())
-    state_en += transcription + "+"
+    state_en += transcription + " "
     vi_text = translate_en2vi(transcription)
-    state_vi += vi_text + "+"
+    state_vi += vi_text + " "
     return state_en, state_vi
 
 """Gradio demo"""

requirements.txt

@@ -1,22 +1,17 @@
-transformers==4.21.1
-transformers[sentencepiece]
-transformers[onnx]
-datasets==2.4.0
-optimum==1.3.0
-sacremoses==0.0.53
-onnxruntime==1.12.1
-onnx==1.12.0
-optimum[graphcore]
-torch
-torchaudio
+torch==1.9.0
+torchaudio==0.9.0
+transformers==4.9.2 
+datasets==1.11.0 
 pyctcdecode==v0.1.0
 speechbrain
 pydub
 kenlm
+pyctcdecode
 soundfile
 ffmpeg-python
 gradio
 nltk
 librosa
-https://github.com/kpu/kenlm/archive/master.zip
-
+transformers
+transformers[sentencepiece]
+https://github.com/kpu/kenlm/archive/master.zip