app.py

import gradio as gr
from transformers import pipeline
import torch
import numpy as np
device = "cuda:0" if torch.cuda.is_available() else "cpu"

wav2_ft = pipeline("automatic-speech-recognition",model='sanchit-gandhi/wav2vec2-large-tedlium',device=device,trust_remote_code=True)


app = gr.Blocks()

def inference(path):
  out = wav2_ft(
      path,
      max_new_tokens=256,
      chunk_length_s=30,
      batch_size=8,
  )
  return out['text']


def transcribe(stream, new_chunk):
    sr, y = new_chunk
    y = y.astype(np.float32)
    y /= np.max(np.abs(y))

    if stream is not None:
        stream = np.concatenate([stream, y])
    else:
        stream = y
    return stream, wav2_ft({"sampling_rate": sr, "raw": stream})["text"] 

mic_mode = gr.Interface(
    fn=inference,
    inputs=gr.Audio(sources="microphone", type='filepath', label="Record Your Lecture"),
    outputs=gr.Textbox(label="Transcription Output"),
    title="🎙️ Recording & Transcribe",
    description="Record through your mic. When you're done, hit stop and wait a moment. Feel free to trim the recording. Then, hit Submit!",
    examples=[],
)


upload_mode = gr.Interface(
    fn=inference,
    inputs=gr.Audio(sources="upload", type='filepath', label="Upload Your Lecture Recording"),
    outputs=gr.Textbox(label="Transcription Output"),
    title="📂 Upload & Transcribe",
    description="Have a recorded lecture? Upload the audio file here, and it'll be transcribed in seconds!",
)

# inspired by Gradio App Real Time Speech Recognition: https://www.gradio.app/guides/real-time-speech-recognition
live_mode = gr.Interface(
    transcribe,
    ["state", gr.Audio(sources=["microphone"], streaming=True)],
    ["state", "text"],
    title="🎤 Live Transcription",
    description="Transcribe your lecture in real-time! Start speaking into your microphone, and watch the transcription appear instantly.",
    live=True,
)



with app:
    gr.Markdown(
        """
        # Lecture Transcription 📝
        
        Welcome to **Lecture Transcription**, the go-to tool for transcribing lectures accurately. Whether you’re attending a live lecture or revisiting a recorded one, this app will ensure you don’t miss a single detail.

        ## How It Works
        - **Recording Mode:** Record the lecture as it happens. When you stop, your transcription will be generated.
        - **Upload Mode:** Upload your pre-recorded lecture audio files, and receive a precise transcription. Supports various audio formats including WAV, MP3, and more.
        - **Live Mode:** That's right, low-latency live transcription!

        ## Optimized for Technical Oration
        Under the hood, this is a Wav2Vec2 model fine-tuned on the TED-Lium dataset. It's well-versed for 
        accurately transcribing technical speech.
  
        
        """
    )
    gr.TabbedInterface(
        [mic_mode, upload_mode,live_mode],
        ["🎙️ Record & Transcribe", "📂 Upload & Transcribe","🎤 Live Transcribe"]
    )


app.launch(debug=True)