The Sound
Classification Challenge

from transformers import AutoFeatureExtractor
from transformers import ASTForAudioClassification
import torchaudio

# Load model and feature extractor
model = ASTForAudioClassification.from_pretrained(
    "MIT/ast-finetuned-audioset-10-10-0.4593"
)
extractor = AutoFeatureExtractor.from_pretrained(
    "MIT/ast-finetuned-audioset-10-10-0.4593"
)

# Load audio (resample to 16kHz)
waveform, sr = torchaudio.load("audio.wav")
if sr != 16000:
    waveform = torchaudio.transforms.Resample(sr, 16000)(waveform)

# Inference
inputs = extractor(waveform.squeeze(), sampling_rate=16000,
                   return_tensors="pt")
outputs = model(**inputs)
probs = outputs.logits.sigmoid()  # multi-label

# Top 5 predictions
top5 = probs[0].topk(5)
for score, idx in zip(top5.values, top5.indices):
    label = model.config.id2label[idx.item()]
    print(f"{label}: {score:.3f}")

import panns_inference
from panns_inference import AudioTagging
import librosa

# Load pretrained PANNs CNN14
model = AudioTagging(
    checkpoint_path=None,  # auto-downloads
    device='cuda'
)

# Load and preprocess audio
audio, sr = librosa.load("audio.wav", sr=32000, mono=True)

# Run inference
clipwise_output, embedding = model.inference(audio[None, :])

# clipwise_output shape: (1, 527) — sigmoid probs
# embedding shape: (1, 2048) — can use for downstream

# Top predictions
import numpy as np
labels = panns_inference.labels
top_k = np.argsort(clipwise_output[0])[::-1][:5]
for idx in top_k:
    print(f"{labels[idx]}: {clipwise_output[0][idx]:.3f}")

import torch
import torchaudio
from BEATs import BEATs, BEATsConfig

# Load BEATs checkpoint
checkpoint = torch.load("BEATs_iter3_plus_AS2M.pt")
cfg = BEATsConfig(checkpoint['cfg'])
model = BEATs(cfg)
model.load_state_dict(checkpoint['model'])
model.eval()

# Load audio
waveform, sr = torchaudio.load("audio.wav")
if sr != 16000:
    waveform = torchaudio.transforms.Resample(sr, 16000)(waveform)

# Extract features + classify
with torch.no_grad():
    padding_mask = torch.zeros(waveform.shape).bool()
    probs = model.extract_features(
        waveform, padding_mask=padding_mask
    )[0]  # (1, time_steps, 527)

# Average over time for clip-level prediction
clip_probs = probs.mean(dim=1).sigmoid()

import torch
import torchaudio
import matplotlib.pyplot as plt

# Load audio
waveform, sr = torchaudio.load("audio.wav")

# Create mel spectrogram transform
mel_transform = torchaudio.transforms.MelSpectrogram(
    sample_rate=sr,
    n_fft=1024,         # FFT window size
    hop_length=320,      # 20ms hop @ 16kHz
    n_mels=128,          # mel frequency bins
    f_min=50,            # min frequency
    f_max=8000,          # max frequency
)

# Convert to log-mel (decibels)
mel_spec = mel_transform(waveform)
log_mel = torchaudio.transforms.AmplitudeToDB()(mel_spec)

# Visualize
plt.figure(figsize=(10, 4))
plt.imshow(log_mel[0], aspect='auto', origin='lower',
           cmap='magma')
plt.colorbar(label='dB')
plt.xlabel('Time frames')
plt.ylabel('Mel bins')
plt.title('Log-Mel Spectrogram')
plt.tight_layout()
plt.savefig('mel_spectrogram.png', dpi=150)