ssl_ft_pron

SSL-FT-PRON: Fine-tuned SSL Models for Automatic Pronunciation Assessment (APA)

A collection of fine-tuned Self-Supervised Learning (SSL) speech models (Wav2Vec2.0, HuBERT, WavLM) for Automatic Pronunciation Assessment (APA).
Three strategies are provided per backbone:

• CTC: ASR-style head trained with CTC
• Freeze: CNN feature extractor frozen; rest is fine-tuned
• General: no CTC head;

Important: This Hub repository is a collection. Each model lives in a subdirectory.
Load with the full sub-path, e.g. haeylee/ssl_ft_pron/wav2vec2/general/02_wav2vec2-large-960h.

---

Model Details

• Developed by: Haeyoung Lee (haeylee)
• Affiliation (paper): Seoul National University, SNU Spoken Language Processing Lab
• Model type: SSL speech encoders fine-tuned for APA (CTC / General / Freeze)
• Language(s): English (evaluated on Speechocean762)
• Finetuned from: See base_model list above

Model Sources

• Code: https://github.com/hy310/ssl_finetuning
• Paper: Analysis of Various Self-Supervised Learning Models for Automatic Pronunciation Assessment (APSIPA ASC 2024)

---

Uses

• Research/prototyping for pronunciation scoring and representation analysis (e.g., PCA on hidden states).
• Feature extraction for downstream APA tasks.

---

Bias, Risks, and Limitations

• Trained/evaluated on Speechocean762 (read English by L2 speakers). Generalization to other languages/speaking styles is not guaranteed.
• APA relies on subjective human scores; apply domain calibration and monitor subgroup performance. Recommendation: Validate on in-domain data; report uncertainty and subgroup metrics.

---

How to Get Started

Load a CTC model (with CTC head)

from transformers import AutoModelForCTC, AutoProcessor

ckpt = "haeylee/ssl_ft_pron/wav2vec2/ctc/01_wav2vec2-large"
model = AutoModelForCTC.from_pretrained(ckpt)
processor = AutoProcessor.from_pretrained(ckpt)

Load a General / Freeze model (no CTC head)

from transformers import AutoProcessor, Wav2Vec2Model, HubertModel, WavLMModel

# Wav2Vec2 (General)
ckpt = "haeylee/ssl_ft_pron/wav2vec2/general/01_wav2vec2-large"
model = Wav2Vec2Model.from_pretrained(ckpt)
processor = AutoProcessor.from_pretrained(ckpt)

# HuBERT (Freeze)
# ckpt = "haeylee/ssl_ft_pron/hubert/freeze/06_hubert-large-ll60k"
# model = HubertModel.from_pretrained(ckpt)
# processor = AutoProcessor.from_pretrained(ckpt)

# WavLM (General)
# ckpt = "haeylee/ssl_ft_pron/wavlm/general/10_wavlm-large"
# model = WavLMModel.from_pretrained(ckpt)
# processor = AutoProcessor.from_pretrained(ckpt)

Summary:

• CTC: AutoModelForCTC.from_pretrained(...)
• General/Freeze: Wav2Vec2Model / HubertModel / WavLMModel .from_pretrained(...)

---

Training Details

Training Data

• Dataset: Speechocean762
• Preprocessing: We used preprocess_dataset.py (see the GitHub repo) to convert raw audio/labels into Hugging Face datasets format.

Expected processed layout:

/your/data/path/speechocean762/
└── preprocess/
    ├── speechocean_train_ds/
    └── speechocean_test_ds/

Training Procedure

Preprocessing

# Adjust paths inside the script or via CLI args
python preprocess_dataset.py \
  --data_root /your/data/path/speechocean762 \
  --out_dir  /your/data/path/speechocean762/preprocess

General (no CTC head)

Loads encoders with Wav2Vec2Model / HubertModel / WavLMModel .from_pretrained(...) and trains a regression head to predict 4 APA scores.

python train/baseline.py \
  --model_name facebook/hubert-xlarge-ls960-ft \
  --batch_size 4 \
  --learning_rate 1e-5 \
  --num_train_epochs 30

Freeze (feature extractor frozen)

Same as General, but freezes the CNN feature extractor.

python train/freeze.py \
  --model_name facebook/hubert-xlarge-ls960-ft \
  --freeze_feature_extractor \
  --batch_size 4 \
  --learning_rate 1e-5 \
  --num_train_epochs 30

CTC (ASR-style head)

Uses AutoModelForCTC.from_pretrained(...) for CTC training.

python train/ctc.py \
  --model_name facebook/wav2vec2-large \
  --batch_size 4 \
  --learning_rate 1e-5 \
  --num_train_epochs 30

Artifacts saved: model.safetensors, trainer_state.json, training_args.bin, logs, and checkpoints (per run: args.json, trainer_args.json).

---

Evaluation

Testing Data, Factors & Metrics

• Test set: Speechocean762 (held-out split prepared by preprocess_dataset.py)
• Factors: Backbone (Wav2Vec2 / HuBERT / WavLM) × strategy (CTC / General / Freeze)
• Metric: pearsonr (Pearson correlation coefficient, PCC) for Accuracy, Fluency, Prosody, and Total.

---

Citation

@inproceedings{lee2024analysis,
  title={Analysis of Various Self-Supervised Learning Models for Automatic Pronunciation Assessment},
  author={Lee, Haeyoung and Kim, Sunhee and Chung, Minhwa},
  booktitle={2024 Asia Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC)},
  pages={1--6},
  year={2024},
  organization={IEEE}
}

---

Authors & Contact

• Author: Haeyoung Lee (haeylee)
• Email: [email protected]
• Issues/Requests: https://github.com/hy310/ssl_finetuning