Upload 6 files

app.py

@@ -0,0 +1,70 @@
+import os
+import torch
+import torchaudio
+import numpy as np
+import gradio as gr
+from transformers import AutoFeatureExtractor, HubertForSequenceClassification
+
+# ==== 1. Cấu hình đường dẫn và thiết bị ====
+MODEL_PATH = "./voice_emotion_checkpoint"  # Thay đổi nếu cần
+DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+# ==== 2. Load feature extractor và model ====
+feature_extractor = AutoFeatureExtractor.from_pretrained(MODEL_PATH)
+model = HubertForSequenceClassification.from_pretrained(MODEL_PATH).to(DEVICE)
+model.eval()
+
+# Nếu bạn có file id2label.json:
+# import json
+# with open(os.path.join(MODEL_PATH, "id2label.json"), "r", encoding="utf-8") as f:
+#     id2label = json.load(f)
+# Ngược lại:
+id2label = {int(k): v for k, v in model.config.id2label.items()}
+
+# ==== 3. Hàm xử lý và dự đoán ====
+def predict_emotion(audio_filepath):
+    # 1) Load file và chuyển về numpy
+    waveform, sr = torchaudio.load(audio_filepath)           # waveform: Tensor[chân âm][time]
+    waveform = waveform.numpy()                              # -> numpy array
+    # 2) Stereo -> mono
+    if waveform.ndim > 1:
+        waveform = np.mean(waveform, axis=0)
+    # 3) Resample về 16 kHz nếu cần
+    target_sr = feature_extractor.sampling_rate
+    if sr != target_sr:
+        waveform = torchaudio.functional.resample(
+            torch.from_numpy(waveform), orig_freq=sr, new_freq=target_sr
+        ).numpy()
+        sr = target_sr
+    # 4) Feature extraction
+    inputs = feature_extractor(
+        waveform,
+        sampling_rate=sr,
+        return_tensors="pt",
+        padding=True
+    )
+    input_values = inputs.input_values.to(DEVICE)
+    # 5) Inference
+    with torch.no_grad():
+        logits = model(input_values).logits.cpu().numpy()[0]
+        probs = torch.softmax(torch.from_numpy(logits), dim=-1).numpy()
+        pred_id = int(np.argmax(probs))
+    # 6) Chuẩn bị output
+    pred_label = id2label[pred_id]
+    label_probs = {id2label[i]: float(probs[i]) for i in range(len(probs))}
+    return pred_label, label_probs
+
+# ==== 4. Xây dựng giao diện Gradio ====
+demo = gr.Interface(
+    fn=predict_emotion,
+    inputs=gr.Audio(type="filepath", label="Upload or Record Audio"),
+    outputs=[
+        gr.Label(num_top_classes=1, label="Predicted Emotion"),
+        gr.Label(num_top_classes=len(id2label), label="All Probabilities"),
+    ],
+    title="Vietnamese Speech Emotion Recognition",
+    description="Upload hoặc record audio, mô hình sẽ dự đoán cảm xúc (angry, happy, sad, …).",
+)
+
+if __name__ == "__main__":
+    demo.launch(server_name="0.0.0.0", server_port=7860, share=False)

requirements.txt

@@ -0,0 +1,8 @@
+torch>=1.12.0
+torchaudio>=0.12.0
+transformers>=4.21.0
+datasets>=2.0.0
+evaluate>=0.4.0
+numpy>=1.21.0
+scikit-learn>=1.0.0
+gradio>=3.0

voice_emotion_checkpoint/config.json

@@ -0,0 +1,92 @@
+{
+  "activation_dropout": 0.1,
+  "apply_spec_augment": true,
+  "architectures": [
+    "HubertForSequenceClassification"
+  ],
+  "attention_dropout": 0.1,
+  "bos_token_id": 1,
+  "classifier_proj_size": 256,
+  "conv_bias": false,
+  "conv_dim": [
+    512,
+    512,
+    512,
+    512,
+    512,
+    512,
+    512
+  ],
+  "conv_kernel": [
+    10,
+    3,
+    3,
+    3,
+    3,
+    2,
+    2
+  ],
+  "conv_pos_batch_norm": false,
+  "conv_stride": [
+    5,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2
+  ],
+  "ctc_loss_reduction": "sum",
+  "ctc_zero_infinity": false,
+  "do_stable_layer_norm": false,
+  "eos_token_id": 2,
+  "feat_extract_activation": "gelu",
+  "feat_extract_dropout": 0.0,
+  "feat_extract_norm": "group",
+  "feat_proj_dropout": 0.1,
+  "feat_proj_layer_norm": true,
+  "final_dropout": 0.1,
+  "gradient_checkpointing": false,
+  "hidden_act": "gelu",
+  "hidden_dropout": 0.1,
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 768,
+  "id2label": {
+    "0": "angry",
+    "1": "fearful",
+    "2": "happy",
+    "3": "neutral",
+    "4": "sad",
+    "5": "surprised"
+  },
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "label2id": {
+    "angry": 0,
+    "fearful": 1,
+    "happy": 2,
+    "neutral": 3,
+    "sad": 4,
+    "surprised": 5
+  },
+  "layer_norm_eps": 1e-05,
+  "layerdrop": 0.1,
+  "mask_feature_length": 10,
+  "mask_feature_min_masks": 0,
+  "mask_feature_prob": 0.0,
+  "mask_time_length": 10,
+  "mask_time_min_masks": 2,
+  "mask_time_prob": 0.05,
+  "model_type": "hubert",
+  "num_attention_heads": 12,
+  "num_conv_pos_embedding_groups": 16,
+  "num_conv_pos_embeddings": 128,
+  "num_feat_extract_layers": 7,
+  "num_hidden_layers": 12,
+  "pad_token_id": 0,
+  "tokenizer_class": "Wav2Vec2CTCTokenizer",
+  "torch_dtype": "float32",
+  "transformers_version": "4.51.3",
+  "use_weighted_layer_sum": false,
+  "vocab_size": 32
+}

voice_emotion_checkpoint/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e2174f5b573a35c131479df49f27896bf8bf00748a09a68388b1011e6986ed56
+size 378306056

voice_emotion_checkpoint/preprocessor_config.json

@@ -0,0 +1,9 @@
+{
+  "do_normalize": true,
+  "feature_extractor_type": "Wav2Vec2FeatureExtractor",
+  "feature_size": 1,
+  "padding_side": "right",
+  "padding_value": 0,
+  "return_attention_mask": false,
+  "sampling_rate": 16000
+}