app.py

import gradio as gr
import torch
import numpy as np
import librosa
from transformers import Wav2Vec2ForSequenceClassification, Wav2Vec2FeatureExtractor

# --------------------------------------------------
# Load Your Fine-Tuned Model for Fluency Prediction
# --------------------------------------------------
# This model was fine-tuned with labels remapped from [3..10] to [0..7].
# Ensure that "Yilin0601/wav2vec2-fluency-checkpoints" is your correct repo.
model = Wav2Vec2ForSequenceClassification.from_pretrained(
    "Yilin0601/wav2vec2-fluency-checkpoints"
)
feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained(
    "Yilin0601/wav2vec2-fluency-checkpoints"
)

# --------------------------------------------------
# Prediction Function
# --------------------------------------------------
def predict(audio):
    if audio is None:
        return "No audio provided."
    
    # Gradio returns audio as (sample_rate, np.array)
    sample_rate, audio_data = audio

    # Ensure audio is in floating-point (librosa requires float32 or float64)
    if audio_data.dtype not in [np.float32, np.float64]:
        audio_data = audio_data.astype(np.float32)

    # Convert stereo to mono if needed
    if len(audio_data.shape) > 1 and audio_data.shape[1] > 1:
        audio_data = np.mean(audio_data, axis=1)

    # Resample to 16 kHz if necessary
    if sample_rate != 16000:
        audio_data = librosa.resample(audio_data, orig_sr=sample_rate, target_sr=16000)

    # Extract features using the feature extractor
    inputs = feature_extractor(
        audio_data,
        sampling_rate=16000,
        return_tensors="pt",
        padding=True
    )

    # Model inference
    model.eval()
    with torch.no_grad():
        logits = model(**inputs).logits

    # The model outputs an 8-class prediction (0..7), corresponding to original fluency scores [3..10]
    pred_class = torch.argmax(logits, dim=-1).item()
    predicted_level = pred_class + 3  # Map back to [3..10]

    return f"Predicted Fluency Level: {predicted_level}"

# --------------------------------------------------
# Gradio Interface
# --------------------------------------------------
iface = gr.Interface(
    fn=predict,
    inputs=gr.Audio(type="numpy", label="Record or Upload Audio"),
    outputs="text",
    title="L2 English Fluency Predictor",
    description=(
        "<p style='font-size:16px;'>"
        "This demo predicts your English fluency level on a scale from 0 to 10. "
        "It uses a fine-tuned <b>facebook/wav2vec2-base-960h</b> model trained on the "
        "<b>DynamicSuperb/L2EnglishAccuracy_speechocean762-Scoring</b> dataset, which contains "
        "745 labeled audio recordings of non-native English speakers. "
        "To get your fluency score, simply record or upload an audio file. "
        "<br><br>"
        "<b>Note:</b> This prediction is for demo purposes and should be interpreted with caution. "
        "</p>"
    ),
    allow_flagging="never"
)

if __name__ == "__main__":
    iface.launch()