app.py

import spaces
import torch
import gradio as gr
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline
import tempfile
import os

# Available models
MODELS = {
    "alakxender/whisper-small-dv-full": "Whisper Small DV Full",
    #"alakxender/whisper-small-dv-mx02": "Whisper Small DV MX02"
}

# Model configuration constants
BATCH_SIZE = 8
FILE_LIMIT_MB = 1000
CHUNK_LENGTH_S = 10
STRIDE_LENGTH_S = [3,2]

# Global variables for device and model management
device = 0 if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
current_model_id = None
current_model = None
current_processor = None
current_pipe = None

# Define optimized generation arguments
def get_generate_kwargs(model, is_short_audio=False):
    """
    Get appropriate generation parameters based on audio length.
    Short audio transcription benefits from different parameters.
    """
    common_kwargs = {
        "max_new_tokens": model.config.max_target_positions-4,
        "num_beams": 4,
        "condition_on_prev_tokens": False,
    }
    
    if is_short_audio:
        # Parameters optimized for short audio:
        return {
            **common_kwargs,
            "compression_ratio_threshold": 1.5,     # Balanced setting to avoid repetition
            "no_speech_threshold": 0.4,             # Higher threshold to reduce hallucinations
            "repetition_penalty": 1.5,              # Add penalty for repeated tokens
            "return_timestamps": True,              # Get timestamps for better segmentation
        }
    else:
        # Parameters for longer audio:
        return {
            **common_kwargs,
            "compression_ratio_threshold": 1.35,    # Standard compression ratio for longer audio
            "repetition_penalty": 1.2,              # Light penalty for repeated tokens
        }

@spaces.GPU
def transcribe(audio_input, model_choice, progress=gr.Progress()):
    global current_model_id, current_model, current_processor, current_pipe, device, torch_dtype
    
    if audio_input is None:
        raise gr.Error("No audio file submitted! Please upload or record an audio file before submitting your request.")
    
    try:
        # Load the selected model if not already loaded or different model selected
        if current_model_id != model_choice or current_model is None:
            progress(0, desc=f"Loading model: {MODELS[model_choice]}")
            print(f"Loading model: {model_choice}")
            
            # Initialize model with memory optimizations
            progress(0.2, desc="Downloading model weights...")
            model = AutoModelForSpeechSeq2Seq.from_pretrained(
                model_choice, 
                torch_dtype=torch_dtype,
                low_cpu_mem_usage=True,
                use_safetensors=True
            )
            
            progress(0.4, desc="Moving model to device...")
            model.to(device)

            # Initialize processor
            progress(0.6, desc="Loading processor...")
            processor = AutoProcessor.from_pretrained(model_choice)

            # Single pipeline initialization with all components
            progress(0.8, desc="Creating pipeline...")
            pipe = pipeline(
                "automatic-speech-recognition",
                model=model,
                tokenizer=processor.tokenizer,
                feature_extractor=processor.feature_extractor,
                chunk_length_s=CHUNK_LENGTH_S,
                stride_length_s=STRIDE_LENGTH_S,
                batch_size=BATCH_SIZE,
                torch_dtype=torch_dtype,
                device=device,
            )

            # IMPORTANT: Fix for forced_decoder_ids error
            progress(0.9, desc="Configuring model...")
            # Remove forced_decoder_ids from the model's generation config
            if hasattr(model.generation_config, 'forced_decoder_ids'):
                print("Removing forced_decoder_ids from generation config")
                model.generation_config.forced_decoder_ids = None

            # Also check if it's in the model config
            if hasattr(model.config, 'forced_decoder_ids'):
                print("Removing forced_decoder_ids from model config")
                delattr(model.config, 'forced_decoder_ids')
            
            # Update global variables
            current_model_id = model_choice
            current_model = model
            current_processor = processor
            current_pipe = pipe
            
            print(f"Model {model_choice} loaded successfully on {device}")
        
        # Start transcription
        progress(0.95, desc="Processing audio...")
        
        # Use the defined generate_kwargs dictionary
        result = current_pipe(
            audio_input,
            generate_kwargs=get_generate_kwargs(current_model)
        )
        
        progress(1.0, desc="Transcription complete!")
        return result["text"]
        
    except Exception as e:
        # More detailed error logging might be helpful here if issues persist
        print(f"Detailed Error: {e}") 
        raise gr.Error(f"Transcription failed: {str(e)}")

# Custom CSS with modern Gradio styling
custom_css = """
.thaana-textbox textarea {
    font-size: 18px !important;
    font-family: 'MV_Faseyha', 'Faruma', 'A_Faruma', 'Noto Sans Thaana', 'MV Boli' !important;
    line-height: 1.8 !important;
    direction: rtl !important;
}
"""

demo = gr.Blocks(css=custom_css)

file_transcribe = gr.Interface(
    fn=transcribe,
    inputs=[
        gr.Audio(sources=["upload", "microphone"], type="filepath", label="Audio file"),
        gr.Dropdown(
            choices=list(MODELS.keys()),
            value=list(MODELS.keys())[0],  # Default to first model
            label="Select Model",
            info="Choose the Whisper model for transcription"
        )
    ],
    outputs=gr.Textbox(
        label="",
        lines=2,
        elem_classes=["thaana-textbox"],
        rtl=True
    ),
    title="Transcribe Dhivehi Audio",
    description=(
        "Upload an audio file or record using your microphone to transcribe. Select your preferred model from the dropdown."
    ),
    flagging_mode="never",
    examples=[
        ["sample.mp3", "alakxender/whisper-small-dv-full"]  
    ],
    api_name=False,
    cache_examples=False
)

with demo:
    gr.TabbedInterface([file_transcribe], ["Audio file"])

demo.queue().launch()