app.py

import torch
import time
import moviepy.editor as mp
import psutil
import gradio as gr
import spaces
from transformers import pipeline
from transformers.pipelines.audio_utils import ffmpeg_read
import base64
import requests



DEFAULT_MODEL_NAME = "distil-whisper/distil-large-v3"
DEFAULT_MODEL_NAME = "openai/whisper-large-v3"

BATCH_SIZE = 8

print('start app')

device = 0 if torch.cuda.is_available() else "cpu"
if device == "cpu":
    DEFAULT_MODEL_NAME = "openai/whisper-tiny"

def load_pipeline(model_name):
    return pipeline(
        task="automatic-speech-recognition",
        model=model_name,
        chunk_length_s=30,
        device=device,
    )

pipe = load_pipeline(DEFAULT_MODEL_NAME)
openai_pipe=load_pipeline("openai/whisper-large-v3")
default_pipe = load_pipeline(DEFAULT_MODEL_NAME)

#pipe = None


from gpustat import GPUStatCollection

def update_gpu_status():
    if torch.cuda.is_available() == False:
        return "No Nvidia Device"
    try:
        gpu_stats = GPUStatCollection.new_query()
        for gpu in gpu_stats:
            # Assuming you want to monitor the first GPU, index 0
            gpu_id = gpu.index
            gpu_name = gpu.name
            gpu_utilization = gpu.utilization
            memory_used = gpu.memory_used
            memory_total = gpu.memory_total
            memory_utilization = (memory_used / memory_total) * 100
            gpu_status=(f"GPU {gpu_id}: {gpu_name}, Utilization: {gpu_utilization}%, Memory Used: {memory_used}MB, Memory Total: {memory_total}MB, Memory Utilization: {memory_utilization:.2f}%")
            return gpu_status

    except Exception as e:
        print(f"Error getting GPU stats: {e}")
        return torch_update_gpu_status()

def torch_update_gpu_status():
    if torch.cuda.is_available():
        gpu_info = torch.cuda.get_device_name(0)
        gpu_memory = torch.cuda.mem_get_info(0)
        total_memory = gpu_memory[1] / (1024 * 1024)
        free_memory=gpu_memory[0] /(1024 *1024)
        used_memory = (gpu_memory[1] - gpu_memory[0]) / (1024 * 1024)
        
        gpu_status = f"GPU: {gpu_info} Free Memory:{free_memory}MB   Total Memory: {total_memory:.2f} MB  Used Memory: {used_memory:.2f} MB"
    else:
        gpu_status = "No GPU available"
    return gpu_status

def update_cpu_status():
    import datetime
    # Get the current time
    current_time = datetime.datetime.now().time()
    # Convert the time to a string
    time_str = current_time.strftime("%H:%M:%S")

    cpu_percent = psutil.cpu_percent()
    cpu_status = f"CPU Usage: {cpu_percent}% {time_str}"
    return cpu_status

@spaces.GPU
def update_status():
    gpu_status = update_gpu_status()
    cpu_status = update_cpu_status()
    sys_status=gpu_status+"\n\n"+cpu_status
    return sys_status

def refresh_status():
    return update_status()


@spaces.GPU
def transcribe(audio_path, model_name):
    print(str(time.time())+'  start transcribe ')
    
    if audio_path is None:
        raise gr.Error("No audio file submitted! Please upload or record an audio file before submitting your request.")
    if model_name is None:
        model_name=DEFAULT_MODEL_NAME
    audio_path=audio_path.strip()
    model_name=model_name.strip()
    
    global pipe
    if model_name != pipe.model.name_or_path:
        print("old model is:"+ pipe.model.name_or_path )
        if model_name=="openai/whisper-large-v3":
            pipe=openai_pipe
            print(str(time.time())+" use openai model " + pipe.model.name_or_path)
        elif model_name==DEFAULT_MODEL_NAME:
            pipe=default_pipe
            print(str(time.time())+" use default model " + pipe.model.name_or_path)
        else:
            print(str(time.time())+'  start load model ' + model_name)
            pipe = load_pipeline(model_name)
            print(str(time.time())+'  finished load model ' + model_name)
    
    start_time = time.time()  # Record the start time
    print(str(time.time())+'  start processing and set recording start time point')
    # Load the audio file and calculate its duration
    audio = mp.AudioFileClip(audio_path)
    audio_duration = audio.duration
    print(str(time.time())+'   start pipe ')
    text = pipe(audio_path, batch_size=BATCH_SIZE, generate_kwargs={"task": "transcribe"}, return_timestamps=True)["text"]
    end_time = time.time()  # Record the end time

    transcription_time = end_time - start_time  # Calculate the transcription time

    # Create the transcription time output with additional information
    transcription_time_output = (
        f"Transcription Time: {transcription_time:.2f} seconds\n"
        f"Audio Duration: {audio_duration:.2f} seconds\n"
        f"Model Used: {model_name}\n"
        f"Device Used: {'GPU' if torch.cuda.is_available() else 'CPU'}"
    )

    print(str(time.time())+'   return transcribe '+ text )
    
    return text, transcription_time_output

@spaces.GPU
def handle_upload_audio(audio_path,model_name,old_transcription=''):
    print('old_trans:' + old_transcription)
    (text,transcription_time_output)=transcribe(audio_path,model_name)
    return text+'\n\n'+old_transcription, transcription_time_output

def handle_base64_audio(base64_data, model_name, old_transcription=''):
    # Decode base64 data and save it as a temporary audio file
    binary_data = base64.b64decode(base64_data)
    audio_path = "temp_audio.wav"
    with open(audio_path, "wb") as f:
        f.write(binary_data)

    # Transcribe the audio file
    (text, transcription_time_output) = transcribe(audio_path, model_name)

    # Remove the temporary audio file
    import os
    os.remove(audio_path)

    return text + '\n\n' + old_transcription, transcription_time_output


graudio=gr.Audio(type="filepath",show_download_button=True)
grmodel_textbox=gr.Textbox(
            label="Model Name",
            value=DEFAULT_MODEL_NAME,
            placeholder="Enter the model name",
            info="Some available models: distil-whisper/distil-large-v3   distil-whisper/distil-medium.en   Systran/faster-distil-whisper-large-v3    Systran/faster-whisper-large-v3    Systran/faster-whisper-medium    openai/whisper-tiny,   openai/whisper-base,   openai/whisper-medium,    openai/whisper-large-v3",
        )
groutputs=[gr.TextArea(label="Transcription",elem_id="transcription_textarea",interactive=True,lines=20,show_copy_button=True), 
           gr.TextArea(label="Transcription Info",interactive=True,show_copy_button=True)]

mf_transcribe = gr.Interface(
    fn=handle_upload_audio,
    inputs=[
        graudio, #"numpy" or filepath
        #gr.Radio(["transcribe", "translate"], label="Task", value="transcribe"),
        grmodel_textbox,
    ],
    outputs=groutputs,
    theme="huggingface",
    title="Whisper Transcription",
    description=(
        "Scroll to Bottom to show system status.  "
        "Transcribe long-form microphone or audio file after uploaded audio! "
        "Notice: the space need some time to get a gpu to run, so there may be a delay "
    ),
    allow_flagging="never",
)


grmodel_textbox_64=gr.Textbox(
            label="Model Name",
            value=DEFAULT_MODEL_NAME,
            placeholder="Enter the model name",
            info="Some available models: distil-whisper/distil-large-v3   distil-whisper/distil-medium.en   Systran/faster-distil-whisper-large-v3    Systran/faster-whisper-large-v3    Systran/faster-whisper-medium    openai/whisper-tiny,   openai/whisper-base,   openai/whisper-medium,    openai/whisper-large-v3",
        )
groutputs_64=[gr.TextArea(label="Transcription 64",elem_id="transcription_textarea_64",interactive=True,lines=20,show_copy_button=True), 
           gr.TextArea(label="Transcription Info 64",interactive=True,show_copy_button=True)]


base_transcribe= gr.Interface(
                fn=handle_base64_audio,
                inputs=[
                    gr.Textbox(label="Base64 Audio Data URL", placeholder="Enter the base64 audio data URL"),
                    grmodel_textbox_64,
                ],
                outputs=groutputs_64,
            )


demo = gr.Blocks()


@spaces.GPU
def onload():
    while True:
        print('onload loop excution')
        time.sleep(2)
    return update_status();
    

with demo:
    tabbed_interface = gr.TabbedInterface(
        [
            mf_transcribe,
            base_transcribe
        ],
        ["Audio", "Base64 Audio"],
    )

    with gr.Row():
        refresh_button = gr.Button("Refresh Status")

    sys_status_output = gr.Textbox(label="System Status", interactive=False)

    # Link the refresh button to the refresh_status function
    refresh_button.click(refresh_status, None, [sys_status_output])

    graudio.stop_recording(handle_upload_audio, inputs=[graudio, grmodel_textbox, groutputs[0]], outputs=groutputs)
    graudio.upload(handle_upload_audio, inputs=[graudio, grmodel_textbox, groutputs[0]], outputs=groutputs)
    # Load the initial status using update_status function
    demo.load(onload, inputs=None, outputs=None, queue=False)

    

# Launch the Gradio app
demo.launch(share=True)

print('launched\n\n')