Deepgram

Deepgram provides speech-to-text, text-to-speech, and an end-to-end Voice Agent API for building spoken interfaces. Real-time work runs over WebSocket streams; pre-recorded audio goes through REST endpoints. You can chain the STT and TTS APIs around your own LLM, or hand the whole loop to the Voice Agent API.

For transcription, Nova-3 covers 60+ languages including a multilingual mode, with a medical variant for English; Nova-2 covers 40+ languages and has English variants tuned for meetings, phone calls, finance, conversational AI, voicemail, video, medical, drive-thru, automotive, and air traffic control. Flux is a streaming model built for conversational turn-taking, available in English and in a multilingual variant for English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian, and Dutch. Transcription options include personal-information redaction, vocabulary customisation, and domain terminology handling.

For synthesis, Aura-2 offers 88 voices across English, Spanish, German, French, Dutch, Italian, and Japanese, including five Spanish voices that switch between English and Spanish mid-utterance; Flux TTS is a newer streaming synthesis model. SDKs exist for Python, JavaScript, Go, .NET, and Java, an MCP server exposes the API to AI clients, and a self-hosted deployment is available.

Features

• Streaming STT: WebSocket transcription with Nova-3 or the conversational Flux model
• Pre-recorded STT: REST transcription for audio files
• Domain models: Nova-2 variants for meetings, phone calls, finance, medical, and other English domains
• Multilingual coverage: 60+ languages on Nova-3 and 10 languages on Flux multilingual
• Text-to-speech: Aura-2 with 88 voices in seven languages, plus streaming Flux TTS
• Voice Agent API: a single API that runs STT, LLM, and TTS as one conversational loop
• Transcript controls: redaction, custom vocabulary, and keyword boosting on legacy models
• SDKs: Python, JavaScript, Go, .NET, and Java
• MCP server: lets AI clients such as Claude call Deepgram directly
• Self-hosted deployment: run the models in your own infrastructure with optional add-ons