app.py

import os
import torch
import gradio as gr
from transformers import pipeline, Wav2Vec2ProcessorWithLM
from pyannote.audio import Pipeline
import whisperx
from functools import partial

from utils import split 
from utils import speech_to_text as stt

os.environ["TOKENIZERS_PARALLELISM"] = "false"

color_map = {"joy": "green","anger": "red","surprise": "yellow","sadness": "blue","fear": "orange","love": "purple",}

# Audio components
whisper_device = "cuda" if torch.cuda.is_available() else "cpu"
whisper = whisperx.load_model("tiny.en", whisper_device)
alignment_model, metadata = whisperx.load_align_model(language_code="en", device=whisper_device)
speaker_segmentation = Pipeline.from_pretrained("pyannote/speaker-diarization@2.1", use_auth_token=os.environ['ENO_TOKEN'])

speech_to_text = partial(
    stt, 
    speaker_segmentation=speaker_segmentation, 
    whisper=whisper, 
    alignment_model=alignment_model, 
    metadata=metadata, 
    whisper_device=whisper_device
    )

# Get Transformer Models
emotion_pipeline = pipeline(
    "text-classification",
    model="bhadresh-savani/distilbert-base-uncased-emotion",
)
summarization_pipeline = pipeline(
    "summarization",
    model="knkarthick/MEETING_SUMMARY",
)

# Apply models to transcripts
def summarize(diarized, summarization_pipeline):
    text = ""
    for d in diarized:
        text += f"\n{d[1]}: {d[0]}"
    return summarization_pipeline(text)[0]["summary_text"]

def sentiment(diarized, emotion_pipeline):
    customer_sentiments = []

    for i in range(0, len(diarized), 2):
        speaker_speech, speaker_id = diarized[i]
        sentences = split(speaker_speech)

        if "Customer" in speaker_id:
            outputs = emotion_pipeline(sentences)
            for idx, (o, t) in enumerate(zip(outputs, sentences)):
                 customer_sentiments.append((t, o["label"]))
    return customer_sentiments

EXAMPLES = [["Customer_Support_Call.wav"]]

with gr.Blocks() as demo:

    with gr.Row():
        with gr.Column():
            audio = gr.Audio(label="Audio file", type="filepath")
            btn = gr.Button("Transcribe and Diarize")

            gr.Markdown("**Call Transcript:**")
            diarized = gr.HighlightedText(label="Call Transcript")
            gr.Markdown("Summarize Speaker")
            sum_btn = gr.Button("Get Summary")
            summary = gr.Textbox(lines=4)
            sentiment_btn = gr.Button("Get Customer Sentiment")
            analyzed = gr.HighlightedText(color_map=color_map)

        with gr.Column():
            gr.Markdown("## Example Files")
            gr.Examples(
                examples=EXAMPLES,
                inputs=[audio],
                outputs=[diarized],
                fn=speech_to_text,
                cache_examples=True
            )
    # when example button is clicked, convert audio file to text and diarize
    btn.click(fn=speech_to_text, inputs=audio, outputs=diarized)
    # when summarize checkboxes are changed, create summary
    sum_btn.click(fn=partial(summarize, summarization_pipeline=summarization_pipeline), inputs=[diarized], outputs=summary)
    # when sentiment button clicked, display highlighted text and plot
    sentiment_btn.click(fn=partial(sentiment, emotion_pipeline=emotion_pipeline), inputs=diarized, outputs=[analyzed])

demo.launch(debug=1)