app.py

import soundfile as sf
import datetime
from pyctcdecode import BeamSearchDecoderCTC
import torch
import os
import time
import gc
import gradio as gr
import librosa
from transformers import Wav2Vec2ForCTC, Wav2Vec2ProcessorWithLM, AutoModelForSeq2SeqLM, AutoTokenizer
from numba import cuda

# load pretrained model
model = Wav2Vec2ForCTC.from_pretrained("facebook/mms-1b-all")
processor = Wav2Vec2ProcessorWithLM.from_pretrained("jlonsako/mms-1b-all-AmhLM")


#Define Functions

#convert time into .sbv format
def format_time(seconds):
    # Convert seconds to hh:mm:ss,ms format
    return str(datetime.timedelta(seconds=seconds)).replace('.', ',')

#Convert Video/Audio into 16K wav file
def preprocessAudio(audioFile):
    os.system(f"ffmpeg -y -i {audioFile.name} -ar 16000 ./audio.wav")

#Transcribe!!!

def Transcribe(file):
    device = "cuda:0" if torch.cuda.is_available() else "cpu"
    start_time = time.time()
    model.load_adapter("amh")

    preprocessAudio(file)
    #os.system(f"ffmpeg -y -i ./July3_2023_Sermon.mov -ar 16000 ./audio.wav")
    block_size = 30

    transcripts = []
    stream = librosa.stream(
        "./audio.wav",
        block_length=block_size,
        frame_length=16000,
        hop_length=16000
    )

    model.to(device)
    print(f"Model loaded to {device}: Entering transcription phase")

    #Code for timestamping
    encoding_start = 0
    sbv_file = open("subtitle.sbv", "w")

    for speech_segment in stream:  
        if len(speech_segment.shape) > 1:
            speech_segment = speech_segment[:,0] + speech_segment[:,1]
        input_values = processor(speech_segment, sampling_rate=16_000, return_tensors="pt").input_values.to(device)
        with torch.no_grad():
            logits = model(input_values).logits
        if len(logits.shape) == 1:
            print("test")
            logits = logits.unsqueeze(0)
        #predicted_ids = torch.argmax(logits, dim=-1)
        transcription = processor.batch_decode(logits.cpu().numpy()).text
        transcripts.append(transcription[0])

        #Generate timestamps
        encoding_end = encoding_start + block_size
        formatted_start = format_time(encoding_start)
        formatted_end = format_time(encoding_end)

        #Write to the .sbv file
        sbv_file.write(f"{formatted_start},{formatted_end}\n")
        sbv_file.write(f"{transcription[0]}\n\n")
        encoding_start = encoding_end

        # Freeing up memory
        del input_values
        del logits
        #del predicted_ids
        del transcription
        torch.cuda.empty_cache()
        gc.collect()

    # Join all transcripts into a single transcript
    transcript = ' '.join(transcripts)
    sbv_file.close()

    end_time = time.time()
    os.system("rm ./audio.wav")
    print(f"The script ran for {end_time - start_time} seconds.")
    return("subtitle.sbv")
    
demo = gr.Interface(fn=Transcribe, inputs=gr.File(), outputs=gr.File())
#with gr.Blocks() as demo:
    #file_output = gr.Textbox()
    #upload_button = gr.UploadButton("Click to Upload a sermon",
    #                                    file_types=["video", "audio"], file_count="multiple")
    #upload_button.upload(Transcribe, upload_button, file_output)
demo.launch()