When speech-to-text hears something else entirely

Friday fun: this one’s for the transcriptions that should not exist.

I spend a lot of time watching what people build with speech-to-text APIs. Most of it is serious. Call centre analytics, voice agents, meeting transcription. The kind of things that justify the budget line.

But every so often someone sends me a transcription that makes me stop and laugh. Not at the accuracy. At the specific, perfect wrongness of it.

Here are my favourites.

The cat who called tech support

Someone tested their voice agent with a cat. The recording was 14 seconds of meowing. Deepgram Nova-3 returned “hello hello help me”. Which is both wrong and somehow exactly right. The cat did sound like it needed help.

The pronunciation lottery

A developer in the Deepgram Discord transcribed someone saying “I’ll have the quiche” and got back “I’ll have the keys”. French loanwords are hard. But the confidence score was 97%, which means the model was extremely sure this person wanted keys instead of eggs. I still think about that.

Whisper systems

One of the funniest patterns is people testing STT on whispered speech. There is a whole category of transcriptions that read like a sleep-talker’s diary. “The muffins are watching” from someone who said “pass the muffins please”. Utterly wrong and beautifully wrong.

Why STT transcriptions sound completely wrong sometimes

These are not bugs. Well, some of them are. But mostly they are the model doing what models do. Guessing from probability. Speech-to-text does not understand what words mean. It knows which sounds map to which text based on training data. When the audio is noisy, accented, whispered, or just a cat, the model reaches for the nearest high-probability match.

For production use you handle this with language hints, keyword boosting, and domain-specific tuning. But if you never look at the raw transcriptions, you miss the funny ones.

FAQ

Why does STT sometimes return completely wrong words?

STT models map audio features to text probabilities. They do not understand semantics. Background noise, accents, homophones, and unusual contexts push the probability distribution toward unexpected tokens. This is why production systems layer in language models, keyword prompts, and domain training.

How do you fix funny transcriptions in production?

Use keyword boosting, set language hints for domain-specific vocabulary, and tune the endpointing sensitivity. For critical use cases, add a second pass with a language model that checks the transcription against expected patterns.

Is perfect STT accuracy achievable?

No. Human speech is ambiguous. Even humans disagree on what was said in noisy recordings. The goal is acceptable accuracy for the use case. 95% for meeting summaries, higher for medical dictation, different tolerances for voice agents where the LLM can correct minor errors.