app.py

# app.py

import gradio as gr
import warnings
import torch
from transformers import WhisperTokenizer, WhisperForConditionalGeneration, WhisperProcessor
import soundfile as sf
import ffmpeg
import os
from huggingface_hub import InferenceClient
from gradio_client import Client, file
import spaces

warnings.filterwarnings("ignore")

# Load tokenizer, model, and processor
tokenizer = WhisperTokenizer.from_pretrained("NbAiLabBeta/nb-whisper-medium")
model = WhisperForConditionalGeneration.from_pretrained("NbAiLabBeta/nb-whisper-medium")
processor = WhisperProcessor.from_pretrained("NbAiLabBeta/nb-whisper-medium")

# Set up device
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
torch_dtype = torch.float32

# Move model to device
model.to(device)

def convert_audio_format(audio_path):
    output_path = "converted_audio.wav"
    ffmpeg.input(audio_path).output(output_path, format='wav', ar='16000').run(overwrite_output=True)
    return output_path

    
@spaces.GPU(duration=120, queue=False)
def transcribe_audio(audio_file, batch_size=4):
    audio_path = convert_audio_format(audio_file)
    audio_input, sample_rate = sf.read(audio_path)
    chunk_size = 16000 * 28  # 28 seconds chunks
    chunks = [audio_input[i:i + chunk_size] for i in range(0, len(audio_input), chunk_size)]

    transcription = ""
    for i in range(0, len(chunks), batch_size):
        batch_chunks = chunks[i:i + batch_size]
        inputs = processor(batch_chunks, sampling_rate=16000, return_tensors="pt", padding=True)
        inputs = inputs.to(device)
        attention_mask = inputs.attention_mask.to(device) if 'attention_mask' in inputs else None
        with torch.no_grad():
            output = model.generate(
                inputs.input_features,
                max_length=1024,
                num_beams=7,
                attention_mask=attention_mask
            )
        transcription += " ".join(processor.batch_decode(output, skip_special_tokens=True)) + " "

    return transcription.strip()

# HTML
banner_html = """
<div style="text-align: center;">
    <img src="https://huggingface.co/spaces/camparchimedes/ola_s-audioshop/raw/main/Olas%20AudioSwitch%20Shop.png" alt="Banner" width="87%" height="auto">
</div>
<div style="text-align: center; margin-top: 20px;">
    <img src="https://huggingface.co/spaces/camparchimedes/ola_s-audioshop/raw/main/picture.jpg" alt="picture" width="50%" height="auto">
</div>
"""

images_path = os.path.dirname(__file__)
IMAGES = [
    [
        {
            "text": "What usual stuff happens in this image? :)",
            "files": [f"{images_path}/500x_picture.png"],
        }
    ]
]

# Gradio interface
iface = gr.Blocks()

with iface:
    gr.HTML(banner_html)
    gr.Markdown("# 𝐍𝐯𝐢𝐝𝐢𝐚 𝐀𝟏𝟎𝟎 👋🏼👾🦾⚡ @{NbAiLab/whisper-norwegian-medium}\nUpload audio file:☕")
    audio_input = gr.Audio(type="filepath")
    batch_size_input = gr.Slider(minimum=1, maximum=16, step=1, label="Batch Size")
    transcription_output = gr.Textbox()
    transcribe_button = gr.Button("Transcribe")

    transcribe_button.click(fn=transcribe_audio, inputs=[audio_input, batch_size_input], outputs=transcription_output)

# Launch interface
iface.launch(share=True, debug=True)