audio_inference.py

import torch
import librosa
import numpy as np
from transformers import Wav2Vec2Processor, Wav2Vec2ForSequenceClassification

# =====================
# CONFIG
# =====================
MODEL_DIR = "exported_audio_model"
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
SR = 16000
MAX_SAMPLES = 8 * SR  # 8 seconds

# =====================
# LOAD MODEL + PROCESSOR (ONCE)
# =====================
processor = Wav2Vec2Processor.from_pretrained(MODEL_DIR)
model = Wav2Vec2ForSequenceClassification.from_pretrained(MODEL_DIR)
model.to(DEVICE)
model.eval()

# =====================
# PREDICT FUNCTION
# =====================
def predict_audio(wav_path):
    # Load audio
    audio, sr = librosa.load(wav_path, sr=SR, mono=True)

    # Truncate if needed
    if len(audio) > MAX_SAMPLES:
        audio = audio[:MAX_SAMPLES]

    # Processor handles padding
    inputs = processor(
        audio,
        sampling_rate=SR,
        return_tensors="pt",
        padding=True,
        return_attention_mask=True
    )

    input_values = inputs.input_values.to(DEVICE)
    attention_mask = inputs.attention_mask.to(DEVICE)

    with torch.no_grad():
        outputs = model(
            input_values=input_values,
            attention_mask=attention_mask
        )

        probs = torch.softmax(outputs.logits, dim=1)[0]
        pred_id = torch.argmax(probs).item()

    label = model.config.id2label[pred_id]
    confidence = probs[pred_id].item() * 100

    return label, round(confidence, 2)