Speech-to-text apps won't publish their training data, but their mistakes quietly reveal whose voices they actually learned from.

This explores what commercial speech recognition systems are trained on: whose voices, which accents and dialects, and what recording conditions. The direct answer is that the corpus has no paper that audits a commercial vendor's training set, because vendors don't publish them. What the corpus does have is evidence about training data that you can read from how these systems fail. That indirect evidence turns out to be more revealing than a data sheet would be.

The clearest case is an audit of five major commercial speech-to-text systems. They made nearly 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 researchers traced the gap to the acoustic model, the part that maps sounds to words, and not to vocabulary. In plain terms, the systems had heard too little African American Vernacular English during training. You can't open the training set, but you can infer its shape from who gets misheard. The same method works for text models. Legal reasoning models do worse on older Supreme Court cases because recent cases dominate their training data Why do language models struggle with historical legal cases?. Across both speech and text, an uneven performance profile is a map of what was over- and under-represented.

There's a complication. Modern speech systems increasingly rest on self-supervised models that learn from huge amounts of unlabeled audio. One study finds that these models learn the physics of how the vocal tract produces sound, not the sound categories of any one language Do speech models learn language-specific sounds or universal physics?. That explains why they transfer well across languages, and it suggests that broad, untranscribed audio does a lot of the foundational work. So the question 'what's in the training data' really has two parts: the large unlabeled audio used for pretraining, and the smaller transcribed sets that decide which accents and speaking styles get recognized well. The dialect gap above most likely sits in the second part.

Recording conditions matter too. In noisy real-world settings, speech recognition still gets 15–30% of words wrong. That error rate is why serious dialogue systems treat the transcript as a probable guess rather than a fact Why do dialogue systems need probabilistic reasoning?. The gap between clean benchmark audio and messy deployment audio is itself a clue that training and test data skew toward clean, read speech. A broader caution applies here as well. Work on text LLMs shows that impressive benchmark scores can partly reflect test material leaking into training data How much of LLM few-shot ability comes from training data?. When training data is opaque, published accuracy numbers can't be taken at face value.

The takeaway: with closed commercial systems, you learn what's in the training data mainly by checking whose speech they fail on. The dialect audit shows that this check can trace failures down to a specific component and a specific data gap, which makes it a usable tool for accountability. If you want vendor-level detail on dataset makeup (hours, languages, sources), this corpus doesn't have it.