Deepgram

Deepgram focuses on speech-to-text with an emphasis on speed and real-time performance. The Nova-2 model delivers competitive accuracy with faster-than-real-time processing for pre-recorded audio. For agents, the API offers two primary modes: pre-recorded transcription (POST audio data or URL) and live streaming transcription (WebSocket). Text-to-speech (Aura) enables voice synthesis from text. Audio intelligence features include summarization, topic detection, sentiment analysis, and intent recognition. The API design prioritizes simplicity: transcription is a single endpoint with configuration via query parameters. Deepgram's pricing model (per audio minute) is straightforward. For agents needing fast turnaround on speech-to-text — voice assistants, real-time captioning, meeting transcription — Deepgram is a strong choice.

Pre-recorded transcription: POST audio to /v1/listen with configuration as query parameters (model, language, punctuate, diarize, smart_format, summarize, etc.). The response is synchronous — results return in the same HTTP response, no polling needed. This is simpler than AssemblyAI's async model for pre-recorded audio. Live streaming uses WebSocket at wss://api.deepgram.com/v1/listen with the same configuration parameters. Text-to-speech: POST text to /v1/speak with voice selection. Audio intelligence features are enabled via query parameters (summarize=v2, topics=v2, intents=v2). The query-parameter configuration model is ergonomic for agents — no request body schema to construct for basic transcription. Response includes word-level timestamps, confidence scores, and paragraphs/sentences when enabled.

Pre-recorded API errors return JSON with err_code and err_msg fields. Common errors: invalid audio format, unsupported language, insufficient funds. The synchronous response model for pre-recorded audio eliminates polling-related failure modes. WebSocket streaming requires connection management: keepalive messages, reconnection logic on disconnect, and proper close frame handling. Rate limits are per-project with generous defaults. Audio format support is broad — Deepgram auto-detects format in most cases. The main reliability consideration for real-time streaming: network interruptions require reconnection and potential loss of audio during the gap. Pre-recorded transcription is fast (often faster than audio duration) and reliable. Error messages are clear and actionable.

API keys are created per-project in the Deepgram console. Keys use Bearer or Token authentication via the Authorization header. Projects provide organizational isolation — keys from one project can't access another project's data. Key scopes include member (usage only) and admin (usage + management). This project-based scoping is more granular than single-key APIs but simpler than role-based systems. No OAuth for third-party integrations. API keys don't expire by default. Usage is tracked per-project. For agents, creating a dedicated project per use case provides clean credential isolation. Text-to-speech and speech-to-text use the same API key. On-premises deployment is available for enterprise needs with sensitive audio data.

Documentation at developers.deepgram.com is well-organized with quickstart guides for each language and use case. The API reference documents every query parameter with descriptions and default values. SDKs are available for Python, JavaScript/TypeScript, Go, .NET, and Rust — maintained by Deepgram. The Python and JavaScript SDKs abstract WebSocket connection management for streaming use cases. The playground in the console allows testing transcription with file upload or microphone input. Blog content covers practical integration patterns. Community Discord is active and responsive. The documentation's main strength: it's focused — Deepgram does speech-to-text and TTS, so the docs aren't diluted across a broad product surface. For agents, the quickstart guides provide a fast path to working integration.