import torch
import numpy as np
from transformers import AutoFeatureExtractor
from transformers.models.qwen2_audio.modeling_qwen2_audio import Qwen2AudioEncoder
from transformers.models.qwen2_audio.configuration_qwen2_audio import Qwen2AudioEncoderConfig

# Load model
model = Qwen2AudioEncoder.from_pretrained("Atotti/AFWhisper")
model = model.to("cuda", dtype=torch.bfloat16)
model.eval()

# Load feature extractor (from Qwen2-Audio)
feature_extractor = AutoFeatureExtractor.from_pretrained("Qwen/Qwen2-Audio-7B")

# Load audio (16kHz, 30s fixed length)
import librosa
audio, sr = librosa.load("audio.wav", sr=16000)

# Pad/trim to 30 seconds
target_len = 16000 * 30
if len(audio) < target_len:
    audio = np.pad(audio, (0, target_len - len(audio)))
else:
    audio = audio[:target_len]

# Extract features
inputs = feature_extractor([audio], sampling_rate=16000, return_tensors="pt")
input_features = inputs.input_features.to("cuda", dtype=torch.bfloat16)

# Encode
with torch.no_grad():
    output = model(input_features=input_features)
    features = output.last_hidden_state  # [1, T, 1280]

print(f"Features shape: {features.shape}")

# Mean pooling for utterance-level embedding
embedding = features.mean(dim=1)  # [1, 1280]

Output

Sequential features: [batch, time_steps, 1280] - 時系列特徴量
Pooled embedding: [batch, 1280] - 発話レベル埋め込み

License

See nvidia/audio-flamingo-3 for license information.