Bao2311/wav2vec2-vi-pronunciation-onnx

🎤 Vietnamese Pronunciation Classifier (ONNX)

Mô hình phân loại phát âm tiếng Việt (Miền Nam) — phân biệt phát âm đúng và phát âm sai cho các phụ âm D/R/GI.

Model Details

Attribute	Value
Base Model	wav2vec2-large-vi-vlsp2020
Format	ONNX (cross-platform)
Task	Binary Classification (correct / wrong)
Language	Vietnamese (Southern dialect)
Sample Rate	16 kHz
Audio Duration	500ms per sample
Validation Accuracy	80.3%
ONNX Opset	14

Classes

Label	Class ID	Description
✅ correct	0	Phát âm đúng
❌ wrong	1	Phát âm sai (nhầm D/R/GI)

Quick Start — ONNX Runtime

pip install onnxruntime librosa numpy

import numpy as np
import librosa
import onnxruntime as ort
from huggingface_hub import hf_hub_download

# 1. Download model
repo_id = "Bao2311/wav2vec2-vi-pronunciation-onnx"
onnx_path = hf_hub_download(repo_id, "phoneme_classifier.onnx")
data_path = hf_hub_download(repo_id, "phoneme_classifier.onnx.data")

# 2. Load ONNX session
session = ort.InferenceSession(onnx_path, providers=["CPUExecutionProvider"])

# 3. Load & preprocess audio (16kHz, 500ms)
audio, _ = librosa.load("your_audio.wav", sr=16000, mono=True)
audio = audio[:8000]  # crop to 500ms = 8000 samples
if len(audio) < 8000:
    audio = np.pad(audio, (0, 8000 - len(audio)))

# 4. Normalize
audio = audio.astype(np.float32)
mean, std = audio.mean(), audio.std()
if std > 0:
    audio = (audio - mean) / std

# 5. Inference
input_values = audio.reshape(1, -1)
logits = session.run(None, {"input_values": input_values})[0][0]

# 6. Softmax → prediction
exp_logits = np.exp(logits - np.max(logits))
probs = exp_logits / exp_logits.sum()
pred = int(np.argmax(probs))

labels = {0: "correct ✅", 1: "wrong ❌"}
print(f"Prediction: {labels[pred]} (confidence: {probs[pred]:.1%})")

Files

File	Size	Description
phoneme_classifier.onnx	~2.5 MB	ONNX model graph
phoneme_classifier.onnx.data	~1.2 GB	Model weights (external data)
config.json	<1 KB	Model configuration
preprocessor_config.json	<1 KB	Feature extractor config

Training Details

• Base model: nguyenvulebinh/wav2vec2-large-vi-vlsp2020
• Epochs: 10
• Batch Size: 4
• Learning Rate: 1e-4
• Optimizer: AdamW
• Loss: CrossEntropyLoss
• Frozen layers: Wav2Vec2 feature extractor + first 8 encoder layers

Limitations

• Chỉ hỗ trợ tiếng Việt (giọng Miền Nam)
• Tối ưu cho audio 500ms, 16kHz mono
• Binary classification (đúng/sai), chưa phân loại chi tiết loại lỗi

License

Apache 2.0