app.py

import spaces
import torch
import gradio as gr
import yt_dlp as youtube_dl
from transformers import pipeline, AutoTokenizer, AutoModelForCausalLM, TextIteratorStreamer
from threading import Thread
import tempfile
import os

MODEL_NAME = "openai/whisper-large-v3-turbo"
BATCH_SIZE = 8
FILE_LIMIT_MB = 1000
YT_LENGTH_LIMIT_S = 3600  # limit to 1 hour YouTube files

device = 0 if torch.cuda.is_available() else "cpu"

# Initialize the transcription pipeline
pipe = pipeline(
    task="automatic-speech-recognition",
    model=MODEL_NAME,
    chunk_length_s=30,
    device=device,
)

# Hugging Face Token for the LLM model
HF_TOKEN = os.getenv("HF_TOKEN")  # Make sure to set this in the environment variables

# Load tokenizer and model for SOAP note generation
tokenizer = AutoTokenizer.from_pretrained("NousResearch/Hermes-3-Llama-3.1-8B")
model = AutoModelForCausalLM.from_pretrained("NousResearch/Hermes-3-Llama-3.1-8B", device_map="auto")

# Prompt for SOAP note generation
sys_prompt = "You are a world class clinical assistant."
task_prompt = """
Convert the following transcribed conversation into a clinical SOAP note.
The text includes dialogue between a physician and a patient. Please clearly distinguish between the physician's and the patient's statements.
Extract and organize the information into the relevant sections of a SOAP note:
- Subjective (symptoms and patient statements),
- Objective (clinical findings and observations, these might be missing if the physician has not conducted a physical exam or has not verbally stated findings),
- Assessment (diagnosis or potential diagnoses, objectively provide a top 5 most likely diagnosis based on just the subjective findings, and use the objective findings if available),
- Plan (treatment and follow-up).
Ensure the note is concise, clear, and accurately reflects the conversation.
"""

# Function to transcribe audio inputs
@spaces.GPU
def transcribe(inputs, task):
    if inputs is None:
        raise gr.Error("No audio file submitted! Please upload or record an audio file before submitting your request.")
    text = pipe(inputs, batch_size=BATCH_SIZE, generate_kwargs={"task": task}, return_timestamps=True)["text"]
    return text

# Function to download audio from YouTube
def download_yt_audio(yt_url, filename):
    info_loader = youtube_dl.YoutubeDL()
    try:
        info = info_loader.extract_info(yt_url, download=False)
    except youtube_dl.utils.DownloadError as err:
        raise gr.Error(str(err))

    file_length_s = sum(x * int(t) for x, t in zip([3600, 60, 1], info["duration_string"].split(":")) if t.isdigit())
    if file_length_s > YT_LENGTH_LIMIT_S:
        raise gr.Error(f"Video too long. Maximum allowed duration is {YT_LENGTH_LIMIT_S / 60} minutes.")

    ydl_opts = {"outtmpl": filename, "format": "worstvideo[ext=mp4]+bestaudio[ext=m4a]/best[ext=mp4]/best"}
    with youtube_dl.YoutubeDL(ydl_opts) as ydl:
        ydl.download([yt_url])

# Function to transcribe YouTube audio
@spaces.GPU
def yt_transcribe(yt_url, task):
    with tempfile.TemporaryDirectory() as tmpdirname:
        filepath = os.path.join(tmpdirname, "video.mp4")
        download_yt_audio(yt_url, filepath)
        with open(filepath, "rb") as f:
            inputs = f.read()
    inputs = pipe.feature_extractor.ffmpeg_read(inputs, pipe.feature_extractor.sampling_rate)
    inputs = {"array": inputs, "sampling_rate": pipe.feature_extractor.sampling_rate}
    text = pipe(inputs, batch_size=BATCH_SIZE, generate_kwargs={"task": task}, return_timestamps=True)["text"]
    return f'<iframe width="500" height="320" src="https://www.youtube.com/embed/{yt_url.split("?v=")[-1]}"> </iframe>', text

# Function to generate SOAP notes using LLM
def generate_soap(transcribed_text):
    prompt = f"{sys_prompt}\n\n{task_prompt}\n{transcribed_text}"
    inputs = tokenizer(prompt, return_tensors="pt").input_ids.to(model.device)
    outputs = model.generate(inputs, max_new_tokens=512)
    return tokenizer.decode(outputs[0], skip_special_tokens=True)

# Gradio Interfaces for different inputs
demo = gr.Blocks(theme=gr.themes.Ocean())

mf_transcribe = gr.Interface(
    fn=transcribe,
    inputs=[gr.Audio(sources="microphone", type="filepath"), gr.Radio(["transcribe", "translate"], label="Task", value="transcribe")],
    outputs="text",
    title="Whisper Large V3 Turbo: Transcribe Audio",
    description="Transcribe long-form microphone or audio inputs."
)

file_transcribe = gr.Interface(
    fn=transcribe,
    inputs=[gr.Audio(sources="upload", type="filepath", label="Audio file"), gr.Radio(["transcribe", "translate"], label="Task", value="transcribe")],
    outputs="text",
    title="Whisper Large V3: Transcribe Audio"
)

yt_transcribe = gr.Interface(
    fn=yt_transcribe,
    inputs=[gr.Textbox(lines=1, placeholder="Paste the URL to a YouTube video here", label="YouTube URL"), gr.Radio(["transcribe", "translate"], label="Task", value="transcribe")],
    outputs=["html", "text"],
    title="Whisper Large V3: Transcribe YouTube"
)

soap_note = gr.Interface(
    fn=generate_soap,
    inputs="text",
    outputs="text",
    title="Generate Clinical SOAP Note",
    description="Convert transcribed conversation to a clinical SOAP note with structured sections (Subjective, Objective, Assessment, Plan)."
)

with demo:
    gr.TabbedInterface([mf_transcribe, file_transcribe, yt_transcribe, soap_note], ["Microphone", "Audio file", "YouTube", "SOAP Note"])

demo.queue().launch(ssr_mode=False)