app.py

import gradio as gr
import whisper
import torch
import os
from pydub import AudioSegment, silence
from faster_whisper import WhisperModel  # Import faster-whisper
import numpy as np
from scipy.io import wavfile

# Mapping of model names to Whisper model sizes
MODELS = {
    "Tiny (Fastest)": "tiny",
    "Base (Faster)": "base",
    "Small (Balanced)": "small",
    "Medium (Accurate)": "medium",
    "Large (Most Accurate)": "large",
    "Faster Whisper Large v3": "Systran/faster-whisper-large-v3"  # Renamed and set as default
}

# Mapping of full language names to language codes
LANGUAGE_NAME_TO_CODE = {
    "Auto Detect": "Auto Detect",
    "English": "en",
    "Chinese": "zh",
    "German": "de",
    "Spanish": "es",
    "Russian": "ru",
    "Korean": "ko",
    "French": "fr",
    "Japanese": "ja",
    "Portuguese": "pt",
    "Turkish": "tr",
    "Polish": "pl",
    "Catalan": "ca",
    "Dutch": "nl",
    "Arabic": "ar",
    "Swedish": "sv",
    "Italian": "it",
    "Indonesian": "id",
    "Hindi": "hi",
    "Finnish": "fi",
    "Vietnamese": "vi",
    "Hebrew": "he",
    "Ukrainian": "uk",
    "Greek": "el",
    "Malay": "ms",
    "Czech": "cs",
    "Romanian": "ro",
    "Danish": "da",
    "Hungarian": "hu",
    "Tamil": "ta",
    "Norwegian": "no",
    "Thai": "th",
    "Urdu": "ur",
    "Croatian": "hr",
    "Bulgarian": "bg",
    "Lithuanian": "lt",
    "Latin": "la",
    "Maori": "mi",
    "Malayalam": "ml",
    "Welsh": "cy",
    "Slovak": "sk",
    "Telugu": "te",
    "Persian": "fa",
    "Latvian": "lv",
    "Bengali": "bn",
    "Serbian": "sr",
    "Azerbaijani": "az",
    "Slovenian": "sl",
    "Kannada": "kn",
    "Estonian": "et",
    "Macedonian": "mk",
    "Breton": "br",
    "Basque": "eu",
    "Icelandic": "is",
    "Armenian": "hy",
    "Nepali": "ne",
    "Mongolian": "mn",
    "Bosnian": "bs",
    "Kazakh": "kk",
    "Albanian": "sq",
    "Swahili": "sw",
    "Galician": "gl",
    "Marathi": "mr",
    "Punjabi": "pa",
    "Sinhala": "si",  # Sinhala support
    "Khmer": "km",
    "Shona": "sn",
    "Yoruba": "yo",
    "Somali": "so",
    "Afrikaans": "af",
    "Occitan": "oc",
    "Georgian": "ka",
    "Belarusian": "be",
    "Tajik": "tg",
    "Sindhi": "sd",
    "Gujarati": "gu",
    "Amharic": "am",
    "Yiddish": "yi",
    "Lao": "lo",
    "Uzbek": "uz",
    "Faroese": "fo",
    "Haitian Creole": "ht",
    "Pashto": "ps",
    "Turkmen": "tk",
    "Nynorsk": "nn",
    "Maltese": "mt",
    "Sanskrit": "sa",
    "Luxembourgish": "lb",
    "Burmese": "my",
    "Tibetan": "bo",
    "Tagalog": "tl",
    "Malagasy": "mg",
    "Assamese": "as",
    "Tatar": "tt",
    "Hawaiian": "haw",
    "Lingala": "ln",
    "Hausa": "ha",
    "Bashkir": "ba",
    "Javanese": "jw",
    "Sundanese": "su",
}

# Reverse mapping of language codes to full language names
CODE_TO_LANGUAGE_NAME = {v: k for k, v in LANGUAGE_NAME_TO_CODE.items()}

def detect_language(audio_file):
    """Detect the language of the audio file."""
    # Define device and compute type for faster-whisper
    device = "cuda" if torch.cuda.is_available() else "cpu"
    compute_type = "float32" if device == "cuda" else "int8"
    
    # Load the faster-whisper model for language detection
    model = WhisperModel(MODELS["Faster Whisper Large v3"], device=device, compute_type=compute_type)
    
    # Convert audio to 16kHz mono for better compatibility
    audio = AudioSegment.from_file(audio_file)
    audio = audio.set_frame_rate(16000).set_channels(1)
    processed_audio_path = "processed_audio.wav"
    audio.export(processed_audio_path, format="wav")
    
    # Detect the language using faster-whisper
    segments, info = model.transcribe(processed_audio_path, task="translate", language=None)
    detected_language_code = info.language
    
    # Get the full language name from the code
    detected_language = CODE_TO_LANGUAGE_NAME.get(detected_language_code, "Unknown Language")
    
    # Clean up processed audio file
    os.remove(processed_audio_path)
    
    return f"Detected Language: {detected_language}"

def remove_silence(audio_file, silence_threshold=-40, min_silence_len=500):
    """
    Remove silence from the audio file using AI-based silence detection.
    
    Args:
        audio_file (str): Path to the input audio file.
        silence_threshold (int): Silence threshold in dB. Default is -40 dB.
        min_silence_len (int): Minimum length of silence to remove in milliseconds. Default is 500 ms.
    
    Returns:
        str: Path to the output audio file with silence removed.
    """
    # Load the audio file
    audio = AudioSegment.from_file(audio_file)
    
    # Detect silent chunks
    silent_chunks = silence.detect_silence(
        audio,
        min_silence_len=min_silence_len,
        silence_thresh=silence_threshold
    )
    
    # Remove silent chunks
    non_silent_audio = AudioSegment.empty()
    start = 0
    for chunk in silent_chunks:
        non_silent_audio += audio[start:chunk[0]]  # Add non-silent part
        start = chunk[1]  # Move to the end of the silent chunk
    non_silent_audio += audio[start:]  # Add the remaining part
    
    # Export the processed audio
    output_path = "silence_removed_audio.wav"
    non_silent_audio.export(output_path, format="wav")
    
    return output_path

def detect_voice_activity(audio_file, threshold=0.02):
    """
    Detect voice activity in the audio file and trim the audio to include only voice segments.
    
    Args:
        audio_file (str): Path to the input audio file.
        threshold (float): Amplitude threshold for voice detection. Default is 0.02.
    
    Returns:
        str: Path to the output audio file with only voice segments.
    """
    # Load the audio file
    sample_rate, data = wavfile.read(audio_file)
    
    # Normalize the audio data
    if data.dtype != np.float32:
        data = data.astype(np.float32) / np.iinfo(data.dtype).max
    
    # Detect voice activity
    voice_segments = []
    is_voice = False
    start = 0
    for i, sample in enumerate(data):
        if abs(sample) > threshold and not is_voice:
            is_voice = True
            start = i
        elif abs(sample) <= threshold and is_voice:
            is_voice = False
            voice_segments.append((start, i))
    
    # If the last segment is voice, add it
    if is_voice:
        voice_segments.append((start, len(data)))
    
    # Trim the audio to include only voice segments
    trimmed_audio = np.array([], dtype=np.float32)
    for segment in voice_segments:
        trimmed_audio = np.concatenate((trimmed_audio, data[segment[0]:segment[1]]))
    
    # Export the trimmed audio
    output_path = "voice_trimmed_audio.wav"
    wavfile.write(output_path, sample_rate, trimmed_audio)
    
    return output_path

def transcribe_audio(audio_file, language="Auto Detect", model_size="Faster Whisper Large v3"):
    """Transcribe the audio file."""
    # Convert audio to 16kHz mono for better compatibility
    audio = AudioSegment.from_file(audio_file)
    audio = audio.set_frame_rate(16000).set_channels(1)
    processed_audio_path = "processed_audio.wav"
    audio.export(processed_audio_path, format="wav")
    
    # Load the appropriate model
    if model_size == "Faster Whisper Large v3":
        # Define device and compute type for faster-whisper
        device = "cuda" if torch.cuda.is_available() else "cpu"
        compute_type = "float32" if device == "cuda" else "int8"
        
        # Use faster-whisper for the Systran model
        model = WhisperModel(MODELS[model_size], device=device, compute_type=compute_type)
        segments, info = model.transcribe(
            processed_audio_path,
            task="transcribe",
            word_timestamps=True,
            repetition_penalty=1.1,
            temperature=[0.0, 0.1, 0.2, 0.3, 0.4, 0.6, 0.8, 1.0],
        )
        transcription = " ".join([segment.text for segment in segments])
        detected_language_code = info.language
        detected_language = CODE_TO_LANGUAGE_NAME.get(detected_language_code, "Unknown Language")
    else:
        # Use the standard Whisper model
        model = whisper.load_model(MODELS[model_size])
        
        # Transcribe the audio
        if language == "Auto Detect":
            result = model.transcribe(processed_audio_path, fp16=False)  # Auto-detect language
            detected_language_code = result.get("language", "unknown")
            detected_language = CODE_TO_LANGUAGE_NAME.get(detected_language_code, "Unknown Language")
        else:
            language_code = LANGUAGE_NAME_TO_CODE.get(language, "en")  # Default to English if not found
            result = model.transcribe(processed_audio_path, language=language_code, fp16=False)
            detected_language = language
        
        transcription = result["text"]
    
    # Clean up processed audio file
    os.remove(processed_audio_path)
    
    # Return transcription and detected language
    return f"Detected Language: {detected_language}\n\nTranscription:\n{transcription}"

# Define the Gradio interface
with gr.Blocks() as demo:
    gr.Markdown("# Audio Transcription and Language Detection")
    
    with gr.Tab("Detect Language"):
        gr.Markdown("Upload an audio file to detect its language.")
        detect_audio_input = gr.Audio(type="filepath", label="Upload Audio File")
        detect_language_output = gr.Textbox(label="Detected Language")
        detect_button = gr.Button("Detect Language")
    
    with gr.Tab("Transcribe Audio"):
        gr.Markdown("Upload an audio file, select a language (or choose 'Auto Detect'), and choose a model for transcription.")
        transcribe_audio_input = gr.Audio(type="filepath", label="Upload Audio File")
        language_dropdown = gr.Dropdown(
            choices=list(LANGUAGE_NAME_TO_CODE.keys()),  # Full language names
            label="Select Language",
            value="Auto Detect"
        )
        model_dropdown = gr.Dropdown(
            choices=list(MODELS.keys()),  # Model options
            label="Select Model",
            value="Faster Whisper Large v3",  # Default to "Faster Whisper Large v3"
            interactive=True  # Allow model selection by default
        )
        transcribe_output = gr.Textbox(label="Transcription and Detected Language")
        transcribe_button = gr.Button("Transcribe Audio")
    
    with gr.Tab("Remove Silence"):
        gr.Markdown("Upload an audio file to remove silence.")
        silence_audio_input = gr.Audio(type="filepath", label="Upload Audio File")
        silence_threshold_slider = gr.Slider(
            minimum=-60, maximum=-20, value=-40, step=1,
            label="Silence Threshold (dB)",
            info="Lower values detect quieter sounds as silence."
        )
        min_silence_len_slider = gr.Slider(
            minimum=100, maximum=2000, value=500, step=100,
            label="Minimum Silence Length (ms)",
            info="Minimum duration of silence to remove."
        )
        silence_output = gr.Audio(label="Processed Audio (Silence Removed)", type="filepath")
        silence_button = gr.Button("Remove Silence")
    
    with gr.Tab("Voice Detection and Trimming"):
        gr.Markdown("Upload two audio files to detect voice activity and trim the audio.")
        voice_audio_input1 = gr.Audio(type="filepath", label="Upload Audio File 1")
        voice_audio_input2 = gr.Audio(type="filepath", label="Upload Audio File 2")
        voice_threshold_slider = gr.Slider(
            minimum=0.01, maximum=0.1, value=0.02, step=0.01,
            label="Voice Detection Threshold",
            info="Higher values detect louder sounds as voice."
        )
        voice_output1 = gr.Audio(label="Trimmed Audio 1", type="filepath")
        voice_output2 = gr.Audio(label="Trimmed Audio 2", type="filepath")
        voice_button = gr.Button("Detect and Trim Voice")
    
    # Link buttons to functions
    detect_button.click(detect_language, inputs=detect_audio_input, outputs=detect_language_output)
    transcribe_button.click(
        transcribe_audio,
        inputs=[transcribe_audio_input, language_dropdown, model_dropdown],
        outputs=transcribe_output
    )
    silence_button.click(
        remove_silence,
        inputs=[silence_audio_input, silence_threshold_slider, min_silence_len_slider],
        outputs=silence_output
    )
    voice_button.click(
        lambda audio1, audio2, threshold: (detect_voice_activity(audio1, threshold), detect_voice_activity(audio2, threshold)),
        inputs=[voice_audio_input1, voice_audio_input2, voice_threshold_slider],
        outputs=[voice_output1, voice_output2]
    )

# Launch the Gradio interface
demo.launch()