README.md

---
tags:
- model_hub_mixin
- pytorch_model_hub_mixin
license: apache-2.0
language:
- en
metrics:
- accuracy
base_model:
- microsoft/wavlm-large
pipeline_tag: audio-classification
---
# WavLM-Large for Speech Flow (Fluency) Classification

# Model Description
This model includes the implementation of speech fluency classification described in Vox-Profile: A Speech Foundation Model Benchmark for Characterizing Diverse Speaker and Speech Traits (https://arxiv.org/pdf/2505.14648)

The model first predicts the speech with a 3-second window size and 1-second step size in 
```
["fluent", "disfluent"]
```
If the disfluent speech is detected, we predict the disfluent types in: 
```
[
  "Block", 
  "Prolongation", 
  "Sound Repetition", 
  "Word Repetition", 
  "Interjection"
]
```

# How to use this model

## Download repo
```bash
git clone [email protected]:tiantiaf0627/vox-profile-release.git
```
## Install the package
```bash
conda create -n vox_profile python=3.8
cd vox-profile-release
pip install -e .
```

## Load the model
```python
# Load libraries
import torch
import torch.nn.functional as F
from src.model.fluency.wavlm_fluency import WavLMWrapper

# Find device
device = torch.device("cuda") if torch.cuda.is_available() else "cpu"

# Load model from Huggingface
model = WavLMWrapper.from_pretrained("tiantiaf/wavlm-large-speech-flow").to(device)
model.eval()
```

## Load the model into 3s window chunks
```python
# The way we do inference for fluency is different as the training data is 3s, so we need to do some stacking
audio_data = torch.zeros([1, 16000*10]).float().to(device)
audio_segment = (audio_data.shape[1] - 3*16000) // 16000 + 1
if audio_segment < 1: audio_segment = 1
input_audio = list()
for idx in range(audio_segment): input_audio.append(audio_data[0, 16000*idx:16000*idx+3*16000])
input_audio = torch.stack(input_audio, dim=0)
```
## Prediction
```python
fluency_outputs, disfluency_type_outputs = model(input_audio)
fluency_prob   = F.softmax(fluency_outputs, dim=1).detach().cpu().numpy().astype(float).tolist()

disfluency_type_prob = nn.Sigmoid()(disfluency_type_outputs)
# we can set a higher threshold in practice
disfluency_type_predictions = (disfluency_type_prob > 0.7).int().detach().cpu().numpy().tolist()
disfluency_type_prob = disfluency_type_prob.cpu().numpy().astype(float).tolist()
```

## Now lets gather the predictions for the utterance
```python
utterance_fluency_list = list()
utterance_disfluency_list = list()
for audio_idx in range(audio_segment):
  disfluency_type = list()
  if fluency_prob[audio_idx][0] > 0.5: 
      utterance_fluency_list.append("fluent")
  else: 
      # If the prediction is disfluent, then which disfluency type
      utterance_fluency_list.append("disfluent")
      predictions = disfluency_type_predictions[audio_idx]
      for label_idx in range(len(predictions)):
          if predictions[label_idx] == 1:
            disfluency_type.append(disfluency_type_labels[label_idx])
  utterance_disfluency_list.append(disfluency_type)

# Now print how fluent is the utterance
print(utterance_fluency_list)
print(utterance_disfluency_list)
```

## If you have any questions, please contact: Tiantian Feng ([email protected])

## Kindly cite our paper if you are using our model or find it useful in your work
```
@article{feng2025vox,
  title={Vox-Profile: A Speech Foundation Model Benchmark for Characterizing Diverse Speaker and Speech Traits},
  author={Feng, Tiantian and Lee, Jihwan and Xu, Anfeng and Lee, Yoonjeong and Lertpetchpun, Thanathai and Shi, Xuan and Wang, Helin and Thebaud, Thomas and Moro-Velazquez, Laureano and Byrd, Dani and others},
  journal={arXiv preprint arXiv:2505.14648},
  year={2025}
}
```