Do some voice-to-text apps mishear certain accents or dialects far more often than others, and why?
Do commercial ASR systems show systematic bias across speaker groups?
This explores whether commercial speech-to-text systems mishear some groups of speakers more than others, and if so, where in the system that unevenness comes from.
This explores whether commercial speech-to-text (ASR) systems treat all speakers equally, and if not, where the gap comes from. The short answer is yes. The corpus has one direct study, and it is a stark one. Five major commercial systems made almost twice as many word errors for Black speakers as for white speakers Why do speech-to-text systems fail more often for Black speakers?. The surprising part is the cause. The gap wasn't mainly about vocabulary or slang. It came from the acoustic models, the part of the system that maps sound to speech units, which had seen too little African American Vernacular English during training. So the bias sits at the level of how the system hears, not what words it knows.
That is an odd result when you set it next to a finding about self-supervised speech models. These models seem to learn something close to universal: the physics of how a vocal tract produces sound, rather than the sound categories of any one language Do speech models learn language-specific sounds or universal physics?. If the underlying representations are that general, the doubled error rate suggests the unevenness comes in later. The likely points are the training mix and the tuning that commercial products are built on, rather than anything built into speech modeling itself. The corpus doesn't test this link directly. Still, it gives a useful way to think about the problem: general-purpose foundations don't guarantee even performance once a product is tuned on skewed data.
Text models show the same pattern. Their biases are mostly set during pretraining, and later fine-tuning only nudges them Where do cognitive biases in language models come from?. Patches added at the surface tend to move bias around rather than remove it Can persona prompts actually reduce bias in language models?. For speech recognition, that implies a fix needs better-balanced audio at the foundation, not a post-processing layer. There is also a related point about how common forms win out. Models handle frequent forms best, and users learn to reshape their input toward what the model understands Does high-frequency text homogenize user input before generation?. For speakers of underrepresented dialects, that can mean quietly changing how they talk just to be transcribed.
Speaker-group effects also show up on the other side of voice interfaces. Older adults rated an LLM voice agent lower when it used frequent backchannels like "mm-hm" and "right." Age-related slowing in auditory processing made those quick cues feel disruptive Does frequent backchannel from voice agents help or hurt older adults?. So voice systems can fail particular groups both in how they hear and in how they speak.
The corpus is thin here. It has one direct ASR bias study, and it covers race and not accent, gender, age or non-native speech. The rest is adjacent evidence about where bias comes from and how it behaves. If you want to go deeper, the doubled-error-rate study is the place to start. The pretraining and output-level notes explain why that kind of gap is hard to patch away.
Sources 6 notes
Five major commercial speech-to-text systems produced word error rates nearly twice as high for Black speakers as white speakers, traced to acoustic models trained on insufficient African American Vernacular English data rather than vocabulary differences.
Self-supervised speech models learn the language-agnostic physics of how the vocal tract produces acoustics, not language-specific phonetic categories. This explains their multilingual transfer and predicts their downstream task performance better than phonetic probing.
A causal experiment using random-seed variation and cross-tuning showed that models sharing a pretrained backbone exhibit similar bias patterns regardless of finetuning data. Biases are planted during pretraining and merely swayed by instruction tuning.
Across three models, persona conditioning makes models follow trait instructions but fails to eliminate underlying bias. Between-group sentiment gaps persist unchanged, showing prompts operate only at the output level.
Adam's Law shows LLMs flatten distinct prompts at comprehension time as users rephrase toward higher-frequency forms the model handles best. The same distributional property that creates accuracy on common tasks filters out distinctiveness on the input side.
Show all 6 sources
A CHI 2025 study found that frequent backchannel cues from an LLM voice agent reduced older adults' experience ratings. Age-related slowing in central auditory processing made rapid backchannel signals harder to distinguish and integrate into conversation, feeling disruptive rather than engaging.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- When Patients Cut In: Extending Clinical Conversational AI Safety to Interruptions
- The Illusion of Debiasing: Persona Steering Redistributes Rather Than Reduces Bias in LLMs
- Planted in Pretraining, Swayed by Finetuning: A Case Study on the Origins of Cognitive Biases in LLMs
- Self-Supervised Models of Speech Infer Universal Articulatory Kinematics
- Toward Enabling Natural Conversation with Older Adults via the Design of LLM-Powered Voice Agents that Support Interruptions and Backchannels
- When Persona Attributes Improve Population Alignment in Large Language Models
- When "A Helpful Assistant" Is Not Really Helpful: Personas in System Prompts Do Not Improve Performances of Large Language Models
- Racial disparities in automated speech recognition