app.py

import gradio as gr
import whisper
import torch
from pyannote.audio.pipelines.speaker_verification import PretrainedSpeakerEmbedding
from pyannote.audio import Audio
from pyannote.core import Segment
import subprocess
import wave
import numpy as np
from sklearn.cluster import AgglomerativeClustering
import os
import datetime

# Load models
model_size = "medium"
whisper_model = whisper.load_model(model_size)
embedding_model = PretrainedSpeakerEmbedding(
    "speechbrain/spkrec-ecapa-voxceleb",
    device=torch.device("cuda" if torch.cuda.is_available() else "cpu")
)
audio_processor = Audio()

def process_audio(file_path, num_speakers):
    # Convert to WAV if necessary
    if not file_path.endswith(".wav"):
        wav_path = file_path.replace(file_path.split('.')[-1], 'wav')
        subprocess.call(['ffmpeg', '-i', file_path, wav_path, '-y'])
        file_path = wav_path

    # Get audio duration
    with wave.open(file_path, 'r') as f:
        frames = f.getnframes()
        rate = f.getframerate()
        duration = frames / float(rate)

    # Transcribe audio
    result = whisper_model.transcribe(file_path)
    segments = result["segments"]

    # Generate speaker embeddings
    embeddings = np.zeros(shape=(len(segments), 192))
    for i, segment in enumerate(segments):
        start = segment["start"]
        end = min(duration, segment["end"])
        clip = Segment(start, end)
        waveform, _ = audio_processor.crop(file_path, clip)
        embeddings[i] = embedding_model(waveform[None])
    embeddings = np.nan_to_num(embeddings)

    # Perform clustering
    clustering = AgglomerativeClustering(n_clusters=num_speakers).fit(embeddings)
    labels = clustering.labels_
    for i, segment in enumerate(segments):
        segment["speaker"] = f"SPEAKER {labels[i] + 1}"

    # Generate transcript
    transcript = []
    for segment in segments:
        speaker = segment["speaker"]
        start_time = str(datetime.timedelta(seconds=round(segment["start"])))
        text = segment["text"]
        transcript.append(f"{speaker} ({start_time}): {text}")

    # Clean up
    os.remove(file_path)
    return "\n".join(transcript)

# Gradio interface
def diarize(audio_file, num_speakers):
    file_path = "temp_audio.wav"
    with open(file_path, "wb") as f:
        f.write(audio_file.read())
    return process_audio(file_path, num_speakers)

# UI
interface = gr.Interface(
    fn=diarize,
    inputs=[
        gr.Audio(source="upload", type="file", label="Upload Audio File"),
        gr.Number(label="Number of Speakers", value=2, precision=0),
    ],
    outputs=gr.Textbox(label="Transcript"),
    title="Speaker Diarization & Transcription",
    description="Upload an audio file, specify the number of speakers, and get a diarized transcript."
)

if __name__ == "__main__":
    interface.launch()