app.py

#!/usr/bin/env python

import os
import pathlib

import gradio as gr
import librosa
import spaces
import torch
from transformers import KyutaiSpeechToTextForConditionalGeneration, KyutaiSpeechToTextProcessor

device = "cuda" if torch.cuda.is_available() else "cpu"
model_id = "kyutai/stt-2.6b-en-trfs"
model = KyutaiSpeechToTextForConditionalGeneration.from_pretrained(model_id, device_map=device, torch_dtype="auto")
processor = KyutaiSpeechToTextProcessor.from_pretrained(model_id)

SAMPLE_RATE = 24000
MAX_DURATION = int(os.getenv("MAX_DURATION", "60"))
MAX_SAMPLE_SIZE = SAMPLE_RATE * MAX_DURATION


@spaces.GPU
def transcribe(audio_path: str) -> str:
    """Transcribe an English audio file to text.

    Args:
        audio_path (str): The path to the audio file. The audio must contain English speech.

    Returns:
        str: The transcription of the English audio file.
    """
    if not audio_path:
        return ""

    data, _ = librosa.load(audio_path, sr=SAMPLE_RATE)
    if len(data) > MAX_SAMPLE_SIZE:
        data = data[:MAX_SAMPLE_SIZE]
        gr.Info(f"Audio file is too long. Truncating to {MAX_DURATION} seconds.")

    inputs = processor(data)
    inputs.to(device)
    output_tokens = model.generate(**inputs)
    output = processor.batch_decode(output_tokens, skip_special_tokens=True)
    return output[0]


with gr.Blocks(fill_height=False) as demo:
    # Header
    gr.HTML("""
        <div class="header-container">
            <h1 class="header-title">🎙️ Kyutai Speech-to-Text</h1>
            <p class="header-subtitle">Advanced English Audio Transcription powered by AI</p>
        </div>
    """)

    # Info banner
    gr.HTML(f"""
        <div class="info-banner">
            ℹ️ Upload or record audio in English (max {MAX_DURATION} seconds). Supports WAV, MP3, and other common formats.
        </div>
    """)  # noqa: RUF001

    # Main content
    with gr.Group(elem_classes="main-card"):
        # Audio input
        audio = gr.Audio(
            label="🎵 Audio Input",
            type="filepath",
            sources=["upload", "microphone"],
            elem_classes="audio-container",
        )

        # Transcribe button
        transcribe_btn = gr.Button(
            "✨ Transcribe Audio",
            variant="primary",
            size="lg",
            elem_classes="primary-button",
        )

        # Output
        output = gr.Textbox(
            label="📝 Transcription",
            placeholder="Your transcription will appear here...",
            lines=6,
            max_lines=12,
            elem_classes="transcription-output",
        )

    # Examples section
    with gr.Group(elem_classes="examples-container"):
        gr.Markdown("### 💡 Try These Examples")
        gr.Examples(
            examples=sorted(pathlib.Path("assets").glob("*.wav")) if pathlib.Path("assets").exists() else [],
            inputs=audio,
            outputs=output,
            fn=transcribe,
            examples_per_page=5,
        )

    # Footer
    gr.HTML("""
        <div class="footer-container">
            <p>
                Built with <a href="https://huggingface.co/spaces/akhaliq/anycoder" class="footer-link" target="_blank">anycoder</a> •
                Powered by <a href="https://huggingface.co/kyutai/stt-2.6b-en-trfs" class="footer-link" target="_blank">Kyutai STT 2.6B</a>
            </p>
        </div>
    """)

    # Event handlers
    transcribe_btn.click(
        fn=transcribe,
        inputs=audio,
        outputs=output,
        api_name="transcribe",
    )


if __name__ == "__main__":
    # Custom theme for modern, clean design
    theme = gr.themes.Soft(
        primary_hue="blue",
        secondary_hue="slate",
        neutral_hue="slate",
        font=gr.themes.GoogleFont("Inter"),
        text_size="lg",
        spacing_size="md",
        radius_size="lg",
    ).set(
        button_primary_background_fill="*primary_600",
        button_primary_background_fill_hover="*primary_700",
        block_title_text_weight="600",
        block_label_text_weight="500",
    )
    demo.launch(theme=theme, css_paths="style.css", mcp_server=True)