Deepgram

Streaming speech-to-text transcription optimized for voice agent interactions using the Flux model, which implements built-in turn detection and natural interruption handling via WebSocket (WSS) protocol. Processes audio in real-time with ultra-low latency, automatically detecting speaker intent boundaries without explicit silence detection configuration, enabling natural back-and-forth conversation flows in voice applications.

REST API endpoint for transcribing pre-recorded audio files with automatic language detection across 45+ languages using Nova-3 Multilingual model. Processes complete audio files (not streaming) with configurable accuracy tiers (Base, Enhanced, Nova-1/2, Nova-3) and returns structured transcription with high-accuracy timestamps, speaker diarization, and optional smart formatting for readability.

Choice of multiple STT models with different accuracy-latency-cost tradeoffs: Base (lowest cost, acceptable accuracy), Enhanced (higher accuracy, higher cost), Nova-1/2/3 (highest accuracy, highest cost), and Flux (optimized for real-time conversational use). Users select the appropriate model based on their accuracy requirements and budget, with pricing ranging from $0.0058/min (Nova-1/2) to $0.0165/min (Enhanced).

Enterprise-tier capability to train custom STT models on proprietary data, enabling domain-specific accuracy improvements for specialized vocabularies, accents, or audio characteristics. Custom models are trained on customer-provided audio and transcripts, then deployed as dedicated endpoints with pricing negotiated per use case. Requires enterprise contract and minimum data volume.

Enterprise deployment option to run Deepgram models on customer infrastructure (on-premise or private cloud) rather than using the cloud API. Enables organizations to maintain full data privacy and control, with models deployed as containers or binaries on customer hardware. Requires enterprise contract and self-hosted add-on licensing.

Command-line interface providing direct access to Deepgram API functionality with 28 pre-built commands for transcription, analysis, and model management. Includes built-in Model Context Protocol (MCP) server enabling integration with AI coding tools (Claude, etc.), allowing AI assistants to call Deepgram APIs directly. Eliminates need for custom API client code for common operations.

Rate limiting enforced via concurrent connection limits rather than requests-per-second, with different quotas for each API endpoint and pricing tier. STT streaming supports 150 concurrent WSS connections (Free), 225 (Growth); REST API supports 100 concurrent; TTS supports 45-60 concurrent; Audio Intelligence supports 10 concurrent. Enables predictable scaling for applications with variable request patterns.

Four-tier pricing model: Free tier with $200 credit (no expiration), Pay-As-You-Go with per-minute pricing ($0.0058-$0.0165/min for STT depending on model), Growth tier with annual commitment ($4,000+ minimum, up to 20% discount), and Enterprise tier with custom pricing. Enables organizations to start free and scale to enterprise volumes with predictable costs.

Automatic speaker identification and segmentation integrated into the transcription pipeline, labeling which speaker produced each segment of audio without requiring manual speaker enrollment or pre-training. Uses deep learning to distinguish speakers based on acoustic features and returns speaker labels aligned with transcript timestamps, enabling downstream analysis of conversation dynamics.

Post-transcription audio intelligence API that analyzes transcribed content to extract sentiment (positive/negative/neutral) and detect dominant topics discussed. Operates via REST API on transcription output, applying NLP models to identify emotional tone and subject matter without requiring manual annotation or training data.

Configurable vocabulary boosting mechanism that improves transcription accuracy for domain-specific terms, technical jargon, or proper nouns by providing hints to the STT model during inference. Accepts a list of keywords or phrases and increases their likelihood in the output, useful for medical, legal, technical, or industry-specific audio where standard models may misrecognize specialized terminology.

Built-in language identification that automatically detects the language spoken in audio without requiring explicit language specification. Uses acoustic and linguistic features to identify language at the start of transcription, then routes to the appropriate language-specific model (Nova-3 Multilingual supports 45+ languages). Eliminates the need for users to pre-specify language, enabling language-agnostic transcription pipelines.

REST and WebSocket API for converting text input into natural-sounding speech audio across multiple voices and languages. Supports both single text requests and continuous text streaming, generating audio output in real-time or batch mode. Uses neural vocoding to produce high-quality, natural-sounding speech with configurable voice selection and language routing.

Single WebSocket endpoint that orchestrates speech-to-text, language model inference, and text-to-speech in a unified pipeline, eliminating the need to stitch together separate services. Handles audio input, routes to LLM for processing, and returns synthesized speech output in a single connection, reducing latency and operational complexity. Supports business logic integration and external system calls within the agent flow.

Post-transcription text processing that applies formatting rules to improve readability of raw transcripts, including punctuation insertion, capitalization, number formatting, and sentence segmentation. Converts raw word sequences into properly formatted text suitable for display or documentation without manual editing, using rule-based and learned formatting patterns.

Precise timing information for each word or segment in the transcript, enabling synchronization with video/audio playback and accurate seeking. Timestamps are computed during transcription inference and returned with confidence scores, allowing applications to highlight text as audio plays or enable click-to-seek functionality in media players.