“It feels like we're not meeting the criteria": Examining and Mitigating the Cascading Effects of Bias in Automatic Speech Recognition in Spoken Language Interfaces.

Research Background and Problem Statement

What problems or challenges did the authors identify?

• Automatic Speech Recognition (ASR) systems exhibit significant performance disparities across different demographic groups, particularly between Black and White speech, potentially leading to biased recognition results.
• ASR systems are a critical component of automated feedback systems in education, where their accuracy directly impacts the quality of feedback provided to teachers. However, the authors found that these biases could further affect downstream tasks, such as teacher discourse classification, ultimately resulting in feedback for Black teachers that does not accurately reflect their actual performance.
• AI systems are increasingly used in education to provide feedback to teachers and help improve teaching quality. However, potential biases in these systems could exacerbate inequalities among teachers.

Why is this problem important?

• Fairness in education is crucial for promoting teachers' professional development and ensuring high-quality learning outcomes for students. If AI systems exhibit racial bias, it could undermine teachers' trust in these systems and lead to psychological stress and professional burnout.
• In the long term, the use of biased AI technologies could hinder the career growth of minority teachers, further exacerbating racial and socioeconomic inequalities in education.

Research Motivation and Related Work

• Previous studies have demonstrated that ASR systems exhibit biases in performance across dimensions such as race, dialect, and gender. However, these studies have primarily focused on the ASR systems themselves, neglecting their impact on downstream tasks in real-world scenarios such as education.
• The authors aim to fill this gap by comprehensively studying ASR biases and their downstream effects on teacher discourse classification tasks, particularly in real educational settings.

Solution

What methods or solutions did the authors propose?

The authors proposed the following research and technical approaches:

1. Analyzing the sources of ASR bias: Distinguishing whether speech recognition errors arise from linguistic patterns or audio characteristics to identify the primary sources of bias.
2. Investigating downstream classification bias: Examining how ASR biases increase error rates in teacher discourse classification tasks and their differential impact on White and Black teachers.
3. Bias mitigation techniques:
   • Fine-tuning ASR models: Using diverse training data from the domain (including speech data from both Black and White teachers) to fine-tune the ASR model.
   • ASR-enhanced classification training: Providing the classification model with a mix of ASR-generated and human-transcribed data to improve its robustness in handling ASR errors.

Innovations in the proposed solution

• Extending research on racial bias to encompass the entire computational pipeline of speech recognition systems and downstream tasks, complementing and deepening existing studies.
• Proposing a multi-stage approach (fine-tuning ASR and optimizing downstream classifiers) to reduce bias throughout the computational pipeline.
• Utilizing real-world educational speech data for the study, enhancing the practical applicability of the analysis and solutions.

Implementation Steps

1. Data Collection and Preprocessing:

   • Collecting speech data from Black and White teachers in real educational settings.
   • Comparing human transcriptions with ASR transcriptions to calculate error rates (e.g., word error rate and matching error rate).
   • Using propensity score matching to control for additional variables (e.g., gender, age, and audio signal-to-noise ratio) to further verify the impact of race on error rates.

2. Bias Source Analysis:

   • Segmenting audio into phrases (n-grams) to control for linguistic content and analyzing the impact of audio characteristics on ASR performance.
   • Using speech perplexity scores to test whether linguistic components contribute to bias.

3. Bias Mitigation Experiments:

   • Fine-tuning the ASR model: Training and optimizing the model using speech data from both Black and White teachers.
   • Enhancing classifier training: Training the classification model with mixed transcription data (ASR and human transcriptions).

Research Findings

What specific findings were obtained?

• Sources of ASR Bias:

  • Overall, ASR systems exhibited higher transcription accuracy for White teachers compared to Black teachers. Errors for Black teachers were predominantly deletion errors.
  • The primary source of bias was deficiencies in audio characteristics ("how it is said") rather than linguistic content ("what is said").

• Impact on Downstream Classification:

  • Classification accuracy was lower when using ASR transcriptions compared to human transcriptions, with biases predominantly affecting Black teachers. This led to the misclassification of high-quality teaching discourse from Black teachers.

• Effectiveness of Bias Mitigation:

  • Fine-tuning the ASR model significantly reduced the accuracy gap between Black and White teachers, though it did not completely eliminate the bias.
  • The ASR-enhanced classification training method was less effective than fine-tuning but still improved the classifier's robustness in handling speech from Black teachers to some extent.

Advantages Compared to Existing Solutions

• The study proposed a comprehensive multi-stage approach to address bias, improving both ASR performance and the classification model's ability to handle ASR errors.
• The research findings were successfully applied to real-world educational scenarios, demonstrating practical value.

Experimental or Evaluation Results

• The fine-tuned ASR model significantly improved recognition performance for Black teachers, reducing the word error rate by approximately 10%.
• The classification model's accuracy increased when processing fine-tuned ASR transcriptions, with notable improvements in classifying speech content from Black teachers.

Limitations and Future Directions

• Limitations:

  • The experiments only involved a single ASR system (Whisper), which may limit the generalizability of the findings to other systems.
  • The dataset included a limited number of teachers and only represented Black and White racial groups, lacking diversity in ethnicity and gender.
  • The study did not completely resolve ASR racial bias, indicating the need for further research with more diverse data and advanced bias mitigation techniques.

• Future Directions:

  • Expanding the study to include more speech recognition models and application scenarios to validate the generalizability of the findings.
  • Investigating bias in student speech data to address fairness issues comprehensively in educational speech recognition technologies.
  • Developing new bias mitigation methods, such as using generative models to create more realistic minority speech data or employing counterfactual techniques for model training.

Through this research, the authors provide critical guidance for the fair application of technology in educational contexts, particularly in reducing racial bias in AI systems. This work not only advances the optimization of speech recognition technologies but also offers new perspectives for research on fairness in artificial intelligence.