Speech Technology in CSD

3 basic components of artificial intelligence

-Autonomic speech recognition (ASR): Recognize and transcribe speech (listen to voice then determine phonemes) -Natural language processing (NLP): Understand and interpret language (phonemes and acoustics to process it into language) -Provide information/complete task to respond (listen to command and respond back) Note: AI can be supervised once built or unsupervised (AI thinking for itself)

"solutions" to the AI bias issue

1. Address latent bias in AI algorithms proactively before it occurs, not after 2. Regulatory frameworks governing AI and machine learning algorithms must explicitly include monitoring for biases in performance 3. Recognize the challenges posed by the different perspectives on appropriate uses and applications of AI in medicine, and engage all healthcare stakeholders in the design and implementation of AI.

BIAS DURING Adaptive learning

1. An AI algorithm trained to operate fairly in 1 context could learn from disparities in care in a different context and start to produce biased results Example - You could exclude the system from using biological gender as an identifier, but the system could still learn bias from income -- since incomes is still disproportionate between different biological genders. 2. An AI algorithm might simply learn from pervasive, ongoing, and uncorrected biases in the broader healthcare system that lead to disparate care and outcomes. Example - If people from certain communities are already treated unfairly/incorrectly by the healthcare system, then any data that trains the system on how to care for these communities will automatically be biased in its decision making moving forward

Bias during outcome evaluation

1. When the outcomes of interest or the problems chosen to be solved by AI do not reflect the interests of individual patients or the community. Example - The AI system could be asked to find more efficient ways to treat aphasia. The recommended therapy could be 'effective' and 'efficient' as requested but may be rigorous, complex, and may completely ignore the interests of the patient or their specific social-emotional and pragmatic needs. 2. The outcomes of interest to various healthcare stakeholders—patients, clinicians, health system leaders, payers, etc. —vary widely between stakeholders. Example - The AI system could be asked to evaluate lots of data to find ways to be more efficient and save money. If the system is only concerned about saving money, then it could make recommendations that negatively impact workers or patients in the process.

Benefits of AI

AI tools used in therapy Students use AI speech recognition systems (i.e., Google speech-to-text) in educational settings because they have academic benefits of increased independence to gain educational content, improved sociability, and increased engagement in reading (Cassano et al., 2022).

Automatic speech recognition

Acoustic model predicts phonemic and contextual patterns to identify phonemes

Educational and societal disadvantages for AAE child speakers

Decades and centuries of systemic racism School to prison pipeline: -Lack of access to opportunities and resources in the home -Lack of access to opportunities and resources in education It has come to light that recent advancements in everyday technology, such as artificial intelligence (AI) systems, have inherent biases that may compound the already existing disadvantages. (code switching)

3 basic components of artificial intelligence example

For example, AI systems involve the use of Automatic Speech Recognition (ASR) to translate acoustic speech into text or commands to complete a task.

Bias during clinical implementation

We can over or under diagnose 1. Phenomenon of Automation Bias: Treating AI-based predictions as infallible or following them unquestioningly Example - Following a suggestion from AI to discharge a client from services, even when you know based on clinical experience that it is unethical to do so. 2. Phenomenon of Privilege Bias: Disproportionately benefiting individuals who already experience privilege of one sort or another Example - If you train your AI model on a healthcare system where people with more money can pay for better care, then the model will automatically associate higher income with better health care practices without understanding 'why'

Below the surface of fair AI

Where and who the data is coming from (could produce biases)

Latent bias

an algorithm (not random) may incorrectly identify something based on historical data or because of a stereotype that already exists in society.

Bias in speech recognition (looking at acoustic parameter aka phonemes)

favors mainstream American english. struggles understanding: -different dialects -different accents

Educational and societal disadvantages for AAE child speakers: our role

it is important as a field to do our part to bridge the gap to equity for our AAE speakers -Big goal: the technology is equitable -To improve access which can affect health, educational, and quality of life outcomes for minority children.

AAE adult speaker difficulties with AI systems

•Needed to modify their speech for the technology to work •The technology was not designed to understand their dialect •The technology was not developed for people with an accent