app.py

from transformers import Wav2Vec2Processor, HubertForCTC
import torch
import gradio as gr
import librosa
import numpy as np
import os

# Load processor dan model dari Hugging Face
processor = Wav2Vec2Processor.from_pretrained("Athallah11/hubert-javanase-jasper-finetuned")
model = HubertForCTC.from_pretrained("Athallah11/hubert-javanase-jasper-finetuned")
model.eval()

# Preprocess audio: convert ke float32 dan resample ke 16kHz
def preprocess(audio, sr):
    if not np.issubdtype(audio.dtype, np.floating):
        audio = audio.astype(np.float32)
    if sr != 16000:
        audio = librosa.resample(audio, orig_sr=sr, target_sr=16000)
    return audio

# Fungsi transkripsi utama
def transcribe(audio):
    try:
        # Deteksi format input audio
        if isinstance(audio, tuple):
            sr, audio_data = audio
        elif isinstance(audio, str):
            audio_data, sr = librosa.load(audio, sr=None)
            os.remove(audio)  # hapus file sementara yang diunggah
        else:
            return "Format audio tidak dikenali."

        # Cek durasi minimal agar rekaman tidak terlalu pendek
        durasi = len(audio_data) / sr
        if durasi < 1.0:
            return "Audio terlalu pendek, coba rekam minimal 1 detik."

        input_values = preprocess(audio_data, sr)

        # Normalisasi audio agar volume seragam dan mengurangi noise
        max_abs = np.max(np.abs(input_values))
        if max_abs > 0:
            input_values = input_values / max_abs

        # Tokenisasi dan prediksi
        inputs = processor(input_values, return_tensors="pt", sampling_rate=16000, padding=True)
        with torch.no_grad():
            logits = model(**inputs).logits

        predicted_ids = torch.argmax(logits, dim=-1)
        transcription = processor.batch_decode(predicted_ids)[0]

        return transcription.lower()
    
    except Exception as e:
        return f"Terjadi kesalahan: {str(e)}"

# UI Gradio
iface = gr.Interface(
    fn=transcribe,
    inputs=gr.Audio(sources=["microphone", "upload"], type="numpy", label="Rekam atau Unggah Audio"),
    outputs="text",
    title="Jasper - ASR Bahasa Jawa",
    description="Model ASR Bahasa Jawa fine-tuned dengan HuBERT\n\nTips:\n- Pastikan suara direkam dengan jelas.\n- Rekam minimal 1 detik untuk hasil optimal."
)

if __name__ == "__main__":
    iface.launch()