Model Card: AImpower/StutteredSpeechASR

This model is a version of OpenAI's whisper-large-v2 fine-tuned on the AImpower/MandarinStutteredSpeech dataset, a grassroots-collected corpus of Mandarin Chinese speech from people who stutter (PWS).

Model Details

• Base Model: openai/whisper-large-v2
• Language: Mandarin Chinese
• Fine-tuning Dataset: AImpower/MandarinStutteredSpeech
• Fine-tuning Method: The model was fine-tuned using the LoRA adapter (AdaLora) methodology to preserve speech disfluencies in its transcriptions.
• Paper: Our Collective Voices: The Social and Technical Values of a Grassroots Chinese Stuttered Speech Dataset

Model Description

This model is specifically adapted to provide more accurate and authentic transcriptions for Mandarin-speaking PWS.
Standard Automatic Speech Recognition (ASR) models often exhibit "fluency bias," where they "smoothen" out or delete stuttered speech patterns like repetitions and interjections.
This model was fine-tuned on literal transcriptions that intentionally preserve these disfluencies.

The primary goal is to create a more inclusive ASR system that recognizes and respects the natural speech patterns of PWS, reducing deletion errors and improving overall accuracy.

Intended Uses & Limitations

Intended Use

This model is intended for transcribing conversational Mandarin Chinese speech from individuals who stutter. It's particularly useful for:

• Improving accessibility in speech-to-text applications.
• Linguistic research on stuttered speech.
• Developing more inclusive voice-enabled technologies.

Limitations

• Language Specificity: The model is fine-tuned exclusively on Mandarin Chinese and is not intended for other languages.
• Data Specificity: Performance is optimized for speech patterns present in the AImpower/MandarinStutteredSpeech dataset. It may not perform as well on other types of atypical speech or in environments with significant background noise.
• Variability: Stuttering is highly variable. While the model shows significant improvements across severity levels, accuracy may still vary between individuals and contexts.

---

How to Use

You can use the model with the transformers library. Ensure you have torch, transformers, and librosa installed.

from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor
import torch
import librosa

# Load the fine-tuned model and processor
model_path = "AImpower/StutteredSpeechASR"
model = AutoModelForSpeechSeq2Seq.from_pretrained(model_path)
processor = AutoProcessor.from_pretrained(model_path)

device = "cuda" if torch.cuda.is_available() else "cpu"
model.to(device)

# Load an example audio file (replace with your audio file)
audio_input_name = "example_stuttered_speech.wav"
waveform, sampling_rate = librosa.load(audio_input_name, sr=16000)

# Process the audio and generate transcription
input_features = processor(waveform, sampling_rate=sampling_rate, return_tensors="pt").input_features
input_features = input_features.to(device)

predicted_ids = model.generate(input_features)
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]

print(f"Transcription: {transcription}")

---

Training Data

The model was fine-tuned on the AImpower/MandarinStutteredSpeech dataset.
This dataset was created through a community-led, grassroots effort with StammerTalk, an online community for Chinese-speaking PWS.

• Size: The dataset contains nearly 50 hours of speech from 72 adults who stutter.
• Content: It includes both unscripted, spontaneous conversations between two PWS and the dictation of 200 voice commands.
• Transcription: The training was performed on verbatim (literal) transcriptions that include disfluencies such as word repetitions and interjections, which was a deliberate choice by the community to ensure their speech was represented authentically.

Training Procedure

• Data Split: A three-fold cross-validation approach was used, with data split by participant to ensure robustness. Each fold had a roughly 65:10:25 split for train/dev/test sets, with a balanced representation of mild, moderate, and severe stuttering levels. This model card represents the best-performing fold.
• Hyperparameters:
  • Epochs: 3
  • Learning Rate: 0.001
  • Optimizer: AdamW
  • Batch Size: 16
  • Fine-tuning Method: AdaLora
• GPU: Four NVIDIA A100 80G GPUs

---

Evaluation Results

The fine-tuned model demonstrates a substantial improvement in transcription accuracy across all stuttering severity levels compared to the baseline whisper-large-v2 model.
The key metric used is Character Error Rate (CER), evaluated on literal transcriptions to measure the model's ability to preserve disfluencies.

Stuttering Severity	Baseline Whisper CER	Fine-tuned Model CER
Mild	16.34%	5.80%
Moderate	21.72%	9.03%
Severe	49.24%	20.46%

(Results from Figure 3 of the paper)

Notably, the model achieved a significant reduction in deletion errors (DEL), especially for severe speech (from 26.56% to 2.29%), indicating that it is much more effective at preserving repeated words and phrases instead of omitting them.

Citation

If you use this model, please cite the original paper:

@inproceedings{li2025collective,
  author = {Li, Jingjin and Li, Qisheng and Gong, Rong and Wang, Lezhi and Wu, Shaomei},
  title = {Our Collective Voices: The Social and Technical Values of a Grassroots Chinese Stuttered Speech Dataset},
  year = {2025},
  isbn = {9798400714825},
  publisher = {Association for Computing Machinery},
  address = {New York, NY, USA},
  url = {https://doi.org/10.1145/3715275.3732179},
  booktitle = {The 2025 ACM Conference on Fairness, Accountability, and Transparency},
  pages = {2768–2783},
  location = {Athens, Greece},
  series = {FAccT '25}
}