Speech-to-text (STT)

Speech-to-text, also called automatic speech recognition or transcription, takes audio and returns text, often with timestamps. Options include transcription APIs from general AI providers, speech-specialist vendors, and open models such as Whisper that you can run yourself. Anthropic doesn't offer a speech model, so a Claude-based app pairs a separate STT service with Claude for the language work.

The shape of the feature decides the model. A voice note needs one accurate transcript a few seconds after upload. A meeting recorder needs speaker labels and word timestamps. Live captions and voice assistants need streaming results with low latency. Check each model's limits on file size and duration, since long recordings often have to be split.

Raw transcripts are messy: filler words, odd punctuation, misheard product names, and no structure. Most useful features add an LLM step that cleans the transcript, extracts action items, or fills in a form, while keeping the raw text for reference. Accuracy drops with accents, jargon, crosstalk, and poor microphones, so test with recordings like your users'. Audio is personal data, often from people who never signed up for your product, so handle consent, retention, and deletion deliberately.