app.py

import os

huggingface_token = os.getenv('Hugging_Face_Token')

# Set up Hugging Face API token for authenticated requests
from huggingface_hub.hf_api import HfFolder
HfFolder.save_token(huggingface_token)

from huggingface_hub import HfApi
available_pipelines = [p.modelId for p in HfApi().list_models(filter="pyannote-audio-pipeline")]
list(filter(lambda p: p.startswith("pyannote/"), available_pipelines))

from pyannote.audio import Pipeline
pipeline_diar = Pipeline.from_pretrained("pyannote/speaker-diarization-3.1", use_auth_token=True)

import torch
import torchaudio
import base64
from io import BytesIO

from pyannote.core import Annotation

from pydub import AudioSegment

from transformers import pipeline
transcriber = pipeline(model="openai/whisper-large-v2", return_timestamps=False)

import gradio as gr

def transcribe_recording(base64_encoded_audio):
    """
    Transcribes audio data given as base64 encoded byte string into text and yields
    the transcribed text incrementally as it becomes available.

    Parameters:
    - base64_audio (str): Base64 encoded byte string of audio data.

    Yields:
    - str: Transcribed text incrementally as it becomes available.
    """
    # Decode the base64 encoded string
    base64_decoded_audio = base64.b64decode(base64_encoded_audio)
    
    with BytesIO(base64_decoded_audio) as audio_buffer:
        audio_buffer.seek(0)
        # Load audio with pydub for easy slicing
        audio = AudioSegment.from_file(audio_buffer)
        audio.export("out.mp3", format="mp3")

        # Run diarization pipeline
        dia = pipeline_diar("out.mp3")
        assert isinstance(dia, Annotation)
        # Prepare a list to store data
        data = []

        for i, (speech_turn, track, speaker) in enumerate(dia.itertracks(yield_label=True)):
            # Extract start and end times
            start_time, end_time = speech_turn.start, speech_turn.end

            # Extract and transcribe the audio segment
            segment_audio = audio[int(start_time * 1000):int(end_time * 1000)]
            segment_audio.export("out.mp3", format="mp3")
            text = transcriber("out.mp3")['text']

            # Append the data to the list
            data.append({
                "Start Time": start_time,
                "End Time": end_time,
                "Speaker": speaker,
                "Transcription": text
            })

    return data
    
tanscriber = gr.Interface(
    fn=transcribe_recording,
    inputs="text",
    outputs="text",
    title="whisper transcriber",
).launch()