← all posts site ↗

claude doesn't build its own voice models

i asked claude what it uses for text to speech and speech to text and it got confused. couldn't even do a web search properly. so i had my agent dig into it instead and here's what's actually happening.

claude doesn't use in-house audio models at all. both tts and stt are third party and anthropic isn't hiding it they just don't advertise it loudly. the text to speech is elevenlabs. this is confirmed directly on anthropic's own trust center subprocessor page where elevenlabs is listed as a subcontractor specifically for tts on claude for work. someone spotted this back in may 2025 when anthropic added them to the list.

the speech to text is deepgram with an optional nova-3 model path behind a feature flag. claude code's voice dictation streams raw pcm audio to anthropic's own websocket endpoint but on the server side they run deepgram for the actual transcription. they also send up to fifty domain specific keyterms like project names and git branches and recently accessed file names to deepgram for keyword boosting which is part of why it's good with coding vocabulary.

this tracks with my experience. the tts sounds slightly robotic at first then gets better which is classic elevenlabs streaming behavior where quality improves as more context arrives. the stt is fast because deepgram is known for sub three hundred millisecond streaming latency. unlike openai with their native multimodal audio or google with gemini live anthropic hasn't built any in-house audio models. their voice mode is a classic stt to llm to tts pipeline turn based not duplex. they're betting the model doing the thinking matters more than a custom audio architecture. no research paper exists because there's nothing novel to publish they're just routing through third party providers.