AssemblyAI vs Deepgram vs Cartesia: speech APIs compared

If you are building transcription, meeting tools or voice agents, you will probably evaluate AssemblyAI, Deepgram and Cartesia. All three offer speech-to-text and real-time features through APIs, and all three now target voice agents. Their strengths differ: AssemblyAI emphasises transcription accuracy and language coverage, Deepgram combines fast recognition with self-hosting, and Cartesia focuses on real-time speech generation built on a different model architecture.

The short answer

• Choose AssemblyAI for batch and streaming transcription with broad language coverage and extras such as a dictation API.
• Choose Deepgram for real-time recognition with redaction, conversation-aware models for agents and a self-hosting option.
• Choose Cartesia if natural, low-latency speech generation for live conversations is your priority.

Side by side

Transcription quality and features

AssemblyAI focuses on transcripts that arrive ready to use: formatting, word-level timestamps, speaker information, language detection and key-term prompting, plus a dictation API that removes filler words and resolves self-corrections. Deepgram offers a similar feature set with diarisation and smart formatting, and adds automatic redaction of sensitive data such as card and phone numbers. Accuracy claims from all vendors are their own; test with your audio, accents and vocabulary.

Voice agents

All three now sell voice-agent capabilities. Deepgram’s Flux models are designed to understand turn-taking and interruptions, which is often where voice agents feel unnatural. Cartesia’s architecture, based on state space models, is presented as enabling low latency, and its managed agents are billed per minute from prepaid balances. AssemblyAI sells a voice agent API billed per hour with bring-your-own telephony.

Data control

Deepgram is the only one of the three that offers self-hosting, which matters for teams with data residency requirements, though it adds infrastructure work. For the cloud APIs, review retention and training policies for audio you send.

How to decide

1. Collect a test set of real recordings, including difficult audio and specialist terms.
2. Run it through each API on its free credits and compare error rates on the words that matter to you.
3. For agents, measure response delay and how naturally interruptions are handled.
4. Model the monthly cost at your expected volume using each vendor’s billing unit.
5. Check language coverage and data-handling options for the regions you serve.

AssemblyAIAudioDeepgramAudioCartesiaAudioVapiAutomationRetell AIAutomation

Tools to explore

Google AI StudioCodingGroqCodingOpenAI APICodingOpenRouterCodingDeepSeekCodingBland AIAutomation