app.py

from transformers import pipeline
import gradio as gr
from pyannote.core import Annotation
from pydub import AudioSegment
import torchaudio
from pyannote.audio import Pipeline

diarization_pipe = Pipeline.from_pretrained("pyannote/speaker-diarization@2.1",
                                    use_auth_token="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")  # change to "your-username/the-name-you-picked"



# Load the speech-to-text model (Whisper)
asr_pipe = pipeline("automatic-speech-recognition", model="openai/whisper-tiny")

def transcribe_with_diarization(audio_path):
    # Get speaker segments using the diarization model
  diarization_result = diarization_pipe(audio_path)

    # Extract speaker segments and transcribe them using Whisper ASR
  transcripts = []
  for track, segment,speaker in diarization_result.itertracks(yield_label=True):
      
        print(segment)        
        print(speaker)
        
        start_time = track.start
        end_time = track.end
        print(start_time)          
        print(end_time)   
        label = segment  # Extract the label manually
        waveform, sample_rate = torchaudio.load(audio_path, normalize=True)
        start_sample = int(start_time * sample_rate)
        end_sample = int(end_time * sample_rate)
        print(waveform)
        interval_audio = waveform[:,start_sample:end_sample]
            # Export the interval audio as a temporary WAV file
        torchaudio.save("interval_audio.wav", interval_audio,sample_rate)
        transcript = asr_pipe("interval_audio.wav")
        print(transcript)
        transcripts.append(transcript)

    # Combine the transcriptions from all speakers
  text = " ".join(transcripts)
  return text

iface = gr.Interface(
    fn=transcribe_with_diarization,
    inputs=[
        gr.File(label="Audio File"),
        gr.Audio(source="microphone", type="filepath", filetype="mp3")
    ],
    outputs="text",
    title="Whisper Large Hindi with Speaker Diarization",
    description="Real-time demo for Hindi speech recognition using a fine-tuned Whisper large model with speaker diarization.",
)

iface.launch()